milestone 13 discrepancies: redact credentials from urls in logs, metrics and cli output

.gitea/workflows/ci.yml

@@ -107,8 +107,8 @@ jobs:
           npm ci
           npm run build
 
-      # ReleaseSafe because the < 15 MB budget (PLAN §18) is for release
-      # binaries; a Debug build strips to ~25 MB and can never meet it.
+      # ReleaseSafe because the < 15 MiB budget (PLAN §18) is for release
+      # binaries; a Debug build strips to roughly 25 MiB and can never meet it.
       - name: Build static musl executables
         run: zig build cross -Dweb-dist=web/dist -Doptimize=ReleaseSafe
 
@@ -153,7 +153,7 @@ jobs:
             size=$(stat -c %s "$binary.stripped")
             echo "$triple: stripped size $size bytes"
             if [ "$size" -ge "$size_limit" ]; then
-              echo "stripped executable exceeds the 15 MB budget: $binary"
+              echo "stripped executable exceeds the 15 MiB budget: $binary"
               exit 1
             fi
           done

PLAN.md

@@ -76,7 +76,7 @@ Verified: 0.16.0 ships `std.crypto.tls.Client` only. There is no server-side TLS
 
 - **DB is truth. Config file format is ZON** (`std.zon` parse + stringify — typed parsing into config structs, exact round-trip, stdlib-maintained, comments supported). No TOML: a third-party parser plus a hand-written serializer is two failure surfaces in the correctness-critical bootstrap/round-trip path, bought for syntax familiarity.
 - First start: if DB empty and `/etc/nxdns/config.zon` exists, validate → seed DB. Subsequent starts ignore the file.
-- `nxdns export [--out file.zon]` dumps DB state as canonical ZON. `nxdns import <file.zon>` validates + replaces DB contents (`--force` if DB non-empty). Export/import = backup + host migration, **not** upgrades (§3.7).
+- `nxdns export [--out file.zon]` dumps DB state as canonical ZON. `nxdns import <file.zon>` validates + replaces DB contents (`--force` if the DB holds configuration; client rows materialised from traffic do not count, and survive the replacement. First-seen/last-seen are runtime state, not configuration: they follow the address, so an import never restamps a device the DB already knew). Export/import = backup + host migration, **not** upgrades (§3.7).
 - No file watcher, no auto-regeneration.
 
 ### 3.6 Storage Layout (Decision H)
@@ -130,7 +130,7 @@ IPv4 + IPv6 full parity for: client identity, rate limiting, logging, group assi
 
 - `/etc/nxdns/config.zon` — bootstrap (first start only).
 - `/var/lib/nxdns/config.db`, `/var/lib/nxdns/querylog.db`
-- `/var/lib/nxdns/blocklists/*.list|*.wild`
+- `/var/lib/nxdns/blocklists/*.list|*.wild` (plus `*.raw.tmp|*.list.tmp|*.wild.tmp` during a refresh)
 - `/var/log/nxdns/nxdns.log` — only in file output mode; default is stderr → journald.
 
 ### 3.14 Frontend Stack (Decision I)
@@ -521,7 +521,7 @@ Periodic delete of rows older than `retention_days`; scheduled checkpoint/VACUUM
 
 ### 12.2 Validation
 
-At least one upstream; ports in range; cert+key readable if DoH/DoT server enabled; resolver URLs parseable. `nxdns check` runs the validator + probes upstreams.
+At least one upstream; ports in range; resolver URLs parseable. `nxdns check` runs the validator, probes upstreams, and loads the cert+key pair of each enabled DoH/DoT server through the same `CertStore.init` the listeners use.
 
 ### 12.3 Settings Semantics
 
@@ -565,7 +565,7 @@ Requirements: responsive desktop/mobile; route loaders for initial fetch; TanSta
 ## 15. CLI
 
 - `nxdns run` — start the server.
-- `nxdns check` — validate config, probe upstreams, test cert readability; nonzero exit on failure.
+- `nxdns check` — validate config, probe upstreams, load each enabled listener's certificate and verify its key pairs with it; exit 2 on failure, 0 with warnings.
 - `nxdns export [--out file.zon]`
 - `nxdns import <file.zon> [--force]`
 - `nxdns version` — app version, Zig version string, build date, git commit.
@@ -634,8 +634,8 @@ Exit: documented deployment works end-to-end on the Pi 5.
 - Sustained ≥ 100 qps on Raspberry Pi 5.
 - Blocklist lookup p95 < 1 ms.
 - Cached response p95 < 5 ms.
-- Memory with ~1M blocked domains < 100 MB.
-- Stripped static binary < 10 MB per arch (excluding embedded frontend assets; < 15 MB with them).
+- Memory with ~1M blocked domains < 100 MiB.
+- Stripped static binary < 10 MiB per arch (excluding embedded frontend assets; < 15 MiB with them).
 
 ---
 

docs/explanation/configuration-model.md

@@ -59,11 +59,35 @@ time you look at the UI, rather than "my edit was applied and then quietly
 undone next Tuesday".
 
 The emptiness check is a real query over the content tables, not a flag: the
-database counts as configured when any content table has rows, or when the
-default group has been altered. One consequence is worth knowing, because it
-is not obvious: clients are auto-materialised when they first send a query, so
-a server that has answered even one query is "configured" and will ignore a
-seed file placed there afterwards.
+database counts as configured when a content table holds rows, or when the
+default group has been altered. What it measures is intent, not activity, and
+the client table is where those two come apart. Clients are auto-materialised
+when they first send a query — the DNS path writes a row per device it sees —
+and such a row records what the network did, not what an operator decided. It
+carries `hand_edited = 0`, the emptiness check counts only the `hand_edited = 1`
+rows, and `export` omits them. So answering queries never turns an unconfigured
+database into a configured one; naming a device does.
+
+Counting traffic here would have been a quiet trap: a server that resolved one
+name would have declared itself configured and ignored a seed file placed
+afterwards, and the operator would have had no line of output saying why.
+
+The same distinction survives an import. `import` replaces the content in one
+transaction, which empties `clients` along with every other table, so every
+client row is saved before the wipe. The materialised ones are put back after it
+unchanged, and restoring a backup does not make the server forget the devices it
+has met.
+
+An address the imported file names belongs to the file — the operator's
+statement wins over the discovered row — with one carve-out. First-seen and
+last-seen are not configuration: they record when a device was heard from, the
+configuration model has no field for either, and an import is not a query. So
+they follow the address rather than the row. If the database already knew that
+address, its two timestamps are carried onto the new row; only an address the
+database has never seen takes the import's clock. Without that, re-importing a
+backup would stamp every device the operator had bothered to name as though it
+had just arrived — and those are exactly the devices whose history is worth
+something.
 
 ## The round trip
 
@@ -90,11 +114,17 @@ difference is a configuration change, never noise from when the export ran.
 
 `nxdns import` replaces the whole database content in one transaction. Not a
 merge, not a patch: delete every content table in foreign-key-safe order, then
-insert what the file says. A failed import — bad syntax, failed validation, a
-constraint violation halfway through — leaves the database exactly as it was,
-because everything happens inside a single `BEGIN IMMEDIATE`.
+insert what the file says — with the one exception described above. Every client
+row is lifted out first. The materialised ones are put back rather than
+recreated from a file that never held them, and the observed timestamps of an
+address the file *does* name are merged onto its new row. A client the file
+leaves out is gone, history included: nothing puts a `hand_edited = 1` row back.
+A failed import — bad syntax, failed
+validation, a constraint violation halfway through — leaves the database exactly
+as it was, because everything happens inside a single `BEGIN IMMEDIATE`.
 
-Without `--force`, import refuses a database that already has content. That
+Without `--force`, import refuses a database that already holds configuration.
+That
 check runs *inside* the transaction, after the lock is taken, so it cannot be
 raced by a concurrent write. The effect is that a plain `import` can never
 clobber a configured server by accident, and clobbering it deliberately takes

docs/explanation/performance-and-testing.md

@@ -15,8 +15,8 @@ PLAN §18 sets five:
 - sustained ≥ 100 qps on a Raspberry Pi 5;
 - blocklist lookup p95 < 1 ms;
 - cached response p95 < 5 ms;
-- memory with ~1M blocked domains < 100 MB;
-- stripped static binary < 10 MB per arch, < 15 MB with the embedded frontend.
+- memory with ~1M blocked domains < 100 MiB;
+- stripped static binary < 10 MiB per arch, < 15 MiB with the embedded frontend.
 
 They are household-scale numbers, and they are deliberately unambitious. 100
 qps is far more than a house generates; the point of the target is not speed

docs/how-to/back-up-and-restore.md

@@ -88,8 +88,8 @@ before writing to it.
 
 ## Restore over an existing database
 
-`import` refuses a database that already has content, so a plain `import` can
-never clobber a configured server by accident:
+`import` refuses a database that already holds configuration, so a plain
+`import` can never clobber a configured server by accident:
 
 ```sh
 nxdns import /tmp/nxdns-lab/backup.zon --data-dir /tmp/nxdns-lab/data
@@ -112,6 +112,13 @@ imported /tmp/nxdns-lab/backup.zon
 Stop the server first. `import` replaces the whole configuration underneath a
 process that has already read it, and a running server will not notice.
 
+What `--force` does to the client list is worth knowing before you restore an
+old backup. A client the backup does not name is removed, and its first-seen
+and last-seen go with it — restoring a month-old file drops the devices you
+named since. Devices the server discovered from traffic are kept, and a device
+the backup does name keeps the first-seen and last-seen the database already
+held, so a restore does not restamp your whole network as newly arrived.
+
 On a real install that is the systemd unit:
 
 ```sh
@@ -185,8 +192,9 @@ The data directory layout is in
 ## A backup before every upgrade
 
 There is no downgrade path. Schema migrations run forward automatically at
-startup and before `check`, `export` and `import`; nothing walks them back. Take
-an export before installing a new binary — see [upgrade](upgrade.md).
+startup and before `export` and `import`; nothing walks them back, and `nxdns
+check` does not run them at all. Take an export before installing a new binary —
+see [upgrade](upgrade.md).
 
 Every command on this page was executed on this host as written, except the
 `systemctl` block marked **Not verified on this host** above.

docs/how-to/enable-doh-and-dot.md

@@ -75,23 +75,40 @@ nxdns check --data-dir /tmp/nxdns-lab/data --config /tmp/nxdns-lab/etc/config.zo
 
 ```
 checking configuration file /tmp/nxdns-lab/etc/config.zon
-OK https://cloudflare-dns.com/dns-query
+OK upstreams[0] https://cloudflare-dns.com
 OK: no problems found
 ```
 
-`check` reads both endpoints' certificate and key. An unreadable file is a
-failure and exits 2:
+`check` loads both endpoints' certificate and key the same way the listeners do,
+so what passes here will start. An unreadable file is a failure and exits 2:
 
 ```
-FAIL doh_server.cert_path: '/tmp/nxdns-lab/etc/cert.pem' is not readable
-FAIL dot_server.cert_path: '/tmp/nxdns-lab/etc/cert.pem' is not readable
+FAIL doh_server.cert_path: '/tmp/nxdns-lab/etc/cert.pem': certificate file is not readable
+FAIL dot_server.cert_path: '/tmp/nxdns-lab/etc/cert.pem': certificate file is not readable
 ```
 
-A key readable by anyone but its owner is a warning, and does not change the
-exit code, because the service still starts:
+So is a key that does not belong to the certificate, which is the mistake worth
+catching before a restart — the two files are individually valid and only their
+pairing is wrong. mbedTLS writes its own line to stderr as it rejects the pair:
+
+```
+warning(tls_server): mbedtls_pk_check_pair failed: RSA - Key failed to pass the validity check of the library (-16896)
+warning(tls_server): mbedtls_pk_check_pair failed: RSA - Key failed to pass the validity check of the library (-16896)
+checking configuration file /tmp/nxdns-lab/etc/mismatch.zon
+FAIL doh_server.key_path: '/tmp/nxdns-lab/etc/other.pem': private key does not belong to the certificate
+FAIL dot_server.key_path: '/tmp/nxdns-lab/etc/other.pem': private key does not belong to the certificate
+OK upstreams[0] https://cloudflare-dns.com
+```
+
+A key readable by anyone but its owner is a warning instead. It does not change
+the exit code, because the service still starts, and the summary line counts it
+rather than claiming nothing was found:
 
 ```
 WARN doh_server.key_path: '/tmp/nxdns-lab/etc/key.pem' is mode 644; a TLS key must be readable by its owner only
+WARN dot_server.key_path: '/tmp/nxdns-lab/etc/key.pem' is mode 644; a TLS key must be readable by its owner only
+OK upstreams[0] https://cloudflare-dns.com
+OK: no failures found, 2 warnings
 ```
 
 ## 4. Start and confirm the listeners

docs/how-to/install-with-docker.md

@@ -73,13 +73,28 @@ database has nothing to forward to. The log is
 `no configuration file at '/etc/nxdns/config.zon'; using the database as it is`
 followed by `nxdns run failed: NoUsableUpstreams`.
 
-A file that is present but rejected is a different failure and a different exit
+A file that is present but rejected is a different failure with the same exit
 code. No `default` group, no enabled upstream, a syntax error — `run` prints the
-diagnostic and exits **1**, for example `nxdns run failed: MissingDefaultGroup`.
-Both were run here against this image: a seed file whose only group was named
-`other` exited 1, and an empty `/etc/nxdns` exited 2 with `NoUsableUpstreams`.
-Under `restart: unless-stopped` either one is a restart loop, so read the exit
-code from `docker inspect` to tell them apart; see
+diagnostic and exits 2 as well. Both were run here against this image. A seed
+file whose only group was named `other`:
+
+```
+FAIL groups: no group named 'default'; every unknown client is assigned to it
+nxdns run failed: MissingDefaultGroup
+run `nxdns check` to see the configuration in full
+```
+
+and an empty `/etc/nxdns`:
+
+```
+info(config_bootstrap): no configuration file at '/etc/nxdns/config.zon'; using the database as it is
+nxdns run failed: NoUsableUpstreams
+run `nxdns check` to see the configuration in full
+```
+
+Under `restart: unless-stopped` either one is a restart loop, and the exit code
+alone no longer tells them apart: read the lines above the failure, which either
+name the diagnostic in the file or say there was no file at all. See
 [Troubleshoot nxdns](troubleshoot.md).
 
 The container runs as uid 65532, and the mount is read-only, so the container

docs/how-to/install-with-systemd.md

@@ -152,7 +152,7 @@ A good file prints the source it checked, one `OK` line per upstream, and
 
 ```
 checking configuration file /etc/nxdns/config.zon
-OK https://cloudflare-dns.com/dns-query
+OK upstreams[0] https://cloudflare-dns.com
 OK: no problems found
 ```
 

docs/how-to/set-up-admin-authentication.md

@@ -228,7 +228,7 @@ nxdns export --data-dir /tmp/nxdns-lab/data | grep password
         .password_hash = "$argon2id$v=19$m=19456,t=2,p=1$kvlRj1tdGul3MlfbvzLncLKWirNpJRJ3howFA9/ysgg$7elW7PPQ3WXHwI4YOmOpZ/1KNEQo7ZDLRJhnYOPMjqw",
 ```
 
-`--force` is required because the database already has content. See
+`--force` is required because the database already holds configuration. See
 [back up and restore](back-up-and-restore.md).
 
 ## What happens with no password set

docs/how-to/troubleshoot.md

@@ -21,8 +21,9 @@ nxdns run failed: NoUsableUpstreams
 run `nxdns check` to see the configuration in full
 ```
 
-Exit 2 is reserved for a small set of faults `run` raises itself:
-`NoUsableUpstreams`, `BadBindAddress`, `BadRateLimit` and `BadCertificate`.
+Exit 2 means the configuration is wrong and you can fix it. Every subcommand
+uses the same definition, so a file `run` exits 2 on exits 2 from `check` and
+`import` too.
 
 **Diagnosis.**
 
@@ -45,18 +46,18 @@ checked on its first line.
   certificate. `run` names both paths before it exits:
   `doh_server: '<cert>' + '<key>': certificate file is not readable`.
 
-  **`check` does not catch most of this.** It tests only that each file is
-  readable, and warns when the key is readable beyond its owner; it never opens
-  the PEM. Parsing and the key/certificate pairing happen when `run` builds the
-  TLS context, so `check` can print `OK: no problems found` on a configuration
-  `run` then refuses. Reproduced here with a self-signed pair and the key from a
-  second, unrelated pair:
+  `check` catches this without starting a listener. It loads both PEM files and
+  tests the key against the certificate through the same code `run` uses, so it
+  fails on exactly what `run` would fail on. Reproduced here with a self-signed
+  pair and the key from a second, unrelated pair:
 
   ```
   $ nxdns check --config config.zon
+  warning(tls_server): mbedtls_pk_check_pair failed: RSA - Key failed to pass the validity check of the library (-16896)
   checking configuration file config.zon
-  OK https://cloudflare-dns.com/dns-query
-  OK: no problems found                                             # exit 0
+  FAIL doh_server.key_path: 'mismatched-key.pem': private key does not belong to the certificate
+  OK upstreams[0] https://cloudflare-dns.com
+                                                                    # exit 2
 
   $ nxdns run --config config.zon --data-dir ./data
   warning(tls_server): mbedtls_pk_check_pair failed: RSA - Key failed to pass the validity check of the library (-16896)
@@ -64,59 +65,103 @@ checked on its first line.
   nxdns run failed: BadCertificate                                  # exit 2
   ```
 
-  A cert file containing `not a certificate` behaves the same way — `check`
-  exits 0, `run` exits 2 with `certificate PEM could not be parsed`. So a
-  successful `check` means the paths and permissions are right, not that the
-  certificate is usable; the only test of that is starting the service. Fix the
-  path, the ownership, or the pair; see
+  The `warning(tls_server)` line comes from mbedTLS on stderr and can appear
+  before the `checking` line, which is on stdout. A cert file containing
+  `not a certificate` fails the same way, with
+  `FAIL doh_server.cert_path: 'junk.pem': certificate PEM could not be parsed`.
+  An unreadable file reads
+  `FAIL doh_server.cert_path: '<path>': certificate file is not readable`.
+
+  Fix the path, the ownership, or the pair; see
   [Enable DoH and DoT](enable-doh-and-dot.md).
 - `BadRateLimit` — a rate limit or window is zero. `import` refuses such a
   configuration, so this only reaches a database that was edited by hand.
 - `BadBindAddress` — `dns.bind_ipv4` or `dns.bind_ipv6` is not an address of
   that family.
 
-## The service exits with code 1 on a seed file you just wrote
+## A seed file you just wrote is rejected
 
 **Symptom.** A first start against an empty database prints the validation
-problem and stops, but with exit code 1, not 2:
+problem and stops with exit 2:
 
 ```
-groups: no group named 'default'; every unknown client is assigned to it
+FAIL groups: no group named 'default'; every unknown client is assigned to it
 nxdns run failed: MissingDefaultGroup
+run `nxdns check` to see the configuration in full
 ```
 
 A syntax error behaves the same way:
 
 ```
-config: 2:42: error: expected ',' after initializer
+FAIL config: 3:16: error: expected ',' after initializer
 nxdns run failed: ParseZon
+run `nxdns check` to see the configuration in full
 ```
 
 So does a seed file whose upstream list is empty or all disabled:
 
 ```
-upstreams: at least one upstream must be enabled
+FAIL upstreams: at least one upstream must be enabled
 nxdns run failed: NoUpstreams
+run `nxdns check` to see the configuration in full
 ```
 
 `NoUpstreams` from a seed file is not the same fault as `NoUsableUpstreams`
 above: the first is a file `run` refused, the second is a database `run`
-accepted and found empty. Only the second is exit 2.
+accepted and found empty. Both are exit 2.
 
-**Diagnosis.** Run the same file through `check`, which reports it as a
-configuration problem and exits 2:
+**Diagnosis.** Run the same file through `check`, which reports the same
+problems and exits 2:
 
 ```sh
 nxdns check --config /etc/nxdns/config.zon
 ```
 
 **Fix.** Correct the file the diagnostics name and start again. The database is
-still empty after a failed seed, so the next start re-reads the file. Note that
-`nxdns check` and `nxdns import` of the same bad file exit 2 while `nxdns run`
-exits 1 — the exit code differs by command, the diagnostics do not. All three
-commands were run here against a file missing its `default` group, one with a
-syntax error and one with no enabled upstream, and every pair came out that
-way.
+still empty after a failed seed, so the next start re-reads the file. The exit
+code no longer depends on which command read the file: all three of these files
+were run through `run`, `check` and `import` here, and every one of the nine
+combinations exited 2 with the same diagnostic.
+
+## `nxdns check` fails on a server that is running fine
+
+**Symptom.** The service is up and answering, but `nxdns check` on the same
+machine exits 2 with one long line about a write-ahead log:
+
+```
+checking database /var/lib/nxdns/config.db
+FAIL /var/lib/nxdns/config.db: uncheckpointed changes are waiting in /var/lib/nxdns/config.db-wal, and reading without writing would answer from the older settings in the main file; `nxdns run` applies them. A running nxdns normally holds this log, which is the usual reason to see this line.
+```
+
+Nothing is damaged. `check` opens `config.db` immutable so that it can never
+write to it, and an immutable open ignores the write-ahead log. When that log
+holds bytes, the newest settings are in it and the main file holds older ones,
+so `check` refuses rather than grade stale values.
+
+The log holds bytes after a configuration write that has not been checkpointed
+yet, which on a running server means someone changed something through the web
+interface or the API. A server that has only been answering queries has an empty
+`config.db-wal` and `check` reads it normally — so this line comes and goes, and
+its absence is not proof that nothing is running.
+
+**Fix.** Check the exported configuration instead of the live file:
+
+```sh
+nxdns export --data-dir /var/lib/nxdns --out /tmp/current.zon
+nxdns check --config /tmp/current.zon
+```
+
+`export` opens the database read/write and does see the log, so it renders the
+settings that are actually in force. Stopping the service and checking again
+works too: a clean shutdown checkpoints the log away.
+
+> Reproduced here on a scratch data directory rather than `/var/lib/nxdns` —
+> that path is the only substitution in the output above. nxdns was started on
+> unprivileged ports; `config.db-wal` was 0 bytes and `check` exited 0; one
+> `POST /api/blocklists` took it to 8272 bytes and `check` then printed the line
+> above and exited 2; `export` from the same live directory succeeded and its
+> output checked clean; and after a clean shutdown `check --data-dir` exited 0
+> again.
 
 ## Port 53 is already taken
 

docs/how-to/upgrade.md

@@ -108,43 +108,63 @@ configuration.
 
 ```sh
 nxdns version
-nxdns check
+nxdns export --out /tmp/after-upgrade.zon
+nxdns check --config /tmp/after-upgrade.zon
 dig @127.0.0.1 example.com A +short
 ```
 
-`nxdns check` with no `--config` checks the database, which is what you want
-after an upgrade — it names its source on the first line and migrates a
-database that is one schema version behind before checking it:
+The restart in step 3 is what migrated the database, so by now the schema is
+current and the service is answering. Confirming with `nxdns check` alone would
+not work here, and the reason is worth knowing: `check` opens `config.db`
+immutable so it can never write to it, and the migration you just performed is
+sitting in `config.db-wal` waiting to be checkpointed. Rather than read around
+the log and grade older settings, `check` reports it:
 
 ```
-checking database /var/lib/nxdns/config.db
-OK https://cloudflare-dns.com/dns-query
+FAIL /var/lib/nxdns/config.db: uncheckpointed changes are waiting in /var/lib/nxdns/config.db-wal, and reading without writing would answer from the older settings in the main file; `nxdns run` applies them. A running nxdns normally holds this log, which is the usual reason to see this line.
+```
+
+`export` opens the database read/write and does see the log, so exporting and
+then checking the export validates what is actually in force:
+
+```
+checking configuration file /tmp/after-upgrade.zon
+OK upstreams[0] https://cloudflare-dns.com
 OK: no problems found
 ```
 
-> Verified on this host for the first two commands, with `--data-dir` pointing
-> at the scratch data directory instead of `/var/lib/nxdns`:
+> Verified on this host for the first three commands, with `--data-dir`
+> pointing at a scratch data directory instead of `/var/lib/nxdns` — that path
+> is the only difference from the blocks above:
 >
 > ```
 > $ nxdns version
 > nxdns 0.1.0-dev (unknown)
 > zig 0.16.0
-> $ nxdns check --data-dir $SCRATCH/data
-> checking database /…/scratchpad/data/config.db
-> OK https://cloudflare-dns.com/dns-query
-> OK: no problems found
 > ```
 >
-> The database path on the first line is the only difference from the block
-> above. The `dig` line was not run in this round: nothing is listening on
-> 127.0.0.1:53 here, and port 53 needs root.
+> Against a running server whose database had just been migrated and seeded,
+> `nxdns check --data-dir` printed the uncheckpointed-log line above and exited
+> 2, while `nxdns export --out` followed by `nxdns check --config` on the result
+> exited 0 with `OK: no problems found`. The `dig` line was not run in this
+> round: nothing is listening on 127.0.0.1:53 here, and port 53 needs root.
 
 ## What happens to the database
 
-Migrations run at startup, and also before `check`, `export` and `import`, so
-whichever of those you run first performs the upgrade. A fresh database is
-created at the current schema version; an older one is stepped up to it. The
-log line names both versions:
+Migrations run at startup, and also before `export` and `import`, so whichever
+of those you run first performs the upgrade. `nxdns check` is the exception: it
+opens the database immutable and never migrates, so on a database still one
+version behind it reports the mismatch and exits 2 rather than fixing it:
+
+```
+FAIL /var/lib/nxdns/config.db: schema version 0, this nxdns expects 2; `nxdns run` migrates it, `check` will not
+```
+
+That line was reproduced here against a database stamped at version 0; the path
+and the version numbers are what vary.
+
+A fresh database is created at the current schema version; an older one is
+stepped up to it. The log line names both versions:
 
 ```
 info(migrations): config.db migrated from schema version 0 to 2
@@ -194,9 +214,11 @@ nxdns import config-backup.zon --force
 systemctl start nxdns
 ```
 
-`--force` is required here. A plain `import` into a database that already has
-content fails with `import failed: DatabaseNotEmpty` and exits 2, so it cannot
-clobber a configured server by accident.
+`--force` is required here. A plain `import` into a database that already holds
+configuration fails with `import failed: DatabaseNotEmpty` and exits 2, so it
+cannot clobber a configured server by accident. What counts is what an operator
+set: client rows the DNS path materialised from traffic never trigger the
+refusal on their own.
 
 > Verified on this host for the two `nxdns` lines, against a populated scratch
 > data directory:

docs/reference/cli.md

@@ -23,21 +23,24 @@ Serves DNS until SIGINT or SIGTERM.
 | `--config FILE` | Seed configuration file (default `/etc/nxdns/config.zon`). Read only when the database has never been configured. |
 | `--web-dev DIR` | Serve the web interface from DIR instead of the embedded assets, with no cache headers. Development only. |
 
-Seeding failures are not treated as configuration faults here: a seed file that
-is unparseable, oversized or invalid prints its diagnostics and exits 1, not 2.
-See [exit codes](#exit-codes).
+A seed file that is unparseable, oversized or invalid prints its diagnostics and
+exits 2 — the same code `check` and `import` give for the same file. See
+[exit codes](#exit-codes).
 
 ## `check`
 
-Validates the configuration and probes the upstreams. Exit 0 when clean, 2 when
-it found problems. It always reports every problem, not just the first. What it
-checks, in order:
+Validates the configuration and probes the upstreams. Exit 0 when it found no
+failures, 2 when it found one. It always reports every problem, not just the
+first. What it checks, in order:
 
 1. Which source to check (see [source selection](#source-selection)).
 2. Full validation — the same rules `import` enforces.
-3. For each enabled DoH/DoT listener: certificate and key are readable (FAIL if
-   not), and the key's permissions are owner-only (WARN if any group or other
-   bit is set; the WARN does not change the exit code).
+3. For each enabled DoH/DoT listener: both PEM files are read and the key is
+   tested against the certificate, through the same `CertStore.init` the
+   listeners boot with. A file that is missing, unreadable, too large or
+   unparseable, and a key that does not belong to the certificate, are all FAIL.
+   The key's permissions are a separate finding: WARN when any group or other
+   bit is set, which does not change the exit code.
 4. A live probe: one real A query for `example.com` through every enabled
    upstream, driving the same pool and failover machinery the server uses, with
    `upstream.total_timeout_ms` as the per-attempt deadline. A FAIL line names
@@ -50,13 +53,28 @@ checks, in order:
 | `--data-dir DIR` | Data directory to look for `config.db` in. |
 | `--config FILE` | Check this file instead of the database. |
 
+### Failures and warnings
+
+Every finding carries a severity. `FAIL` is a problem that sets exit 2. `WARN`
+is legal configuration that is almost certainly not what was meant — a blocklist
+source no group links to, a TLS key readable beyond its owner — and never
+changes an exit code, because the service starts either way.
+
+The last line is a summary, and it never contradicts the lines above it:
+
+| What was found | Last line | Exit |
+| --- | --- | --- |
+| Nothing | `OK: no problems found` | 0 |
+| Warnings only | `OK: no failures found, 2 warnings` — `1 warning` in the singular | 0 |
+| At least one failure | No summary line; the FAIL lines are the report | 2 |
+
 ### Source selection
 
 - `--config FILE` given explicitly: check that file, nothing else.
 - Otherwise, if `<data-dir>/config.db` exists: check the database — the right
-  default, since the database is the truth on a configured server. Pending
-  schema migrations are applied first, so a `check` immediately after an upgrade
-  works.
+  default, since the database is the truth on a configured server. It is opened
+  immutable, so `check` writes nothing to it; see
+  [what `check` does not do](#what-check-does-not-do).
 - Otherwise, if the default config file path exists: check it.
 - Otherwise: "nothing to check", exit 2.
 
@@ -64,10 +82,42 @@ The first line of output always names which source was checked. A file larger
 than 4 MiB fails with `larger than 4194304 bytes`; a ZON syntax error is
 reported with its line and column.
 
-A missing file is treated differently depending on how the path was reached. An
-explicit `--config FILE` that does not exist is an I/O failure: `check failed:
-FileNotFound` on stderr, exit 1. A default path that does not exist is one of
-the cases above and produces "nothing to check", exit 2.
+A named file that is missing or unreadable is a finding like any other, not an
+I/O failure that escapes the run: `FAIL <path>: no such file` or
+`FAIL <path>: not readable`, exit 2. That is the same code the implicit path
+gives when there is nothing to check, so naming the file does not change what a
+missing file costs.
+
+### What `check` does not do
+
+`check` never writes to `config.db`. It opens the file immutable, which means
+SQLite refuses every statement that would write and builds no write-ahead log,
+so no `config.db-wal` and no `config.db-shm` appear beside it. It does not chmod
+the file and it does not migrate the schema.
+
+Two consequences are worth knowing before you read a FAIL line as damage:
+
+- A database behind this binary's schema is reported, not upgraded. Start the
+  service to migrate it.
+
+  ```
+  FAIL <db>: schema version <n>, this nxdns expects <m>; `nxdns run` migrates it, `check` will not
+  ```
+
+- A database with an unapplied write-ahead log cannot be graded without writing,
+  because the newest settings are in the log and the main file holds older ones.
+  `check` says so rather than reading the stale values:
+
+  ```
+  FAIL <db>: uncheckpointed changes are waiting in <db>-wal, and reading without writing would answer from the older settings in the main file; `nxdns run` applies them. A running nxdns normally holds this log, which is the usual reason to see this line.
+  ```
+
+  A running nxdns is the usual holder of that log, so this is a common answer
+  when checking a live server rather than a sign of damage. It depends on what
+  is in the log, not on whether a server is up: a configuration write that has
+  not been checkpointed puts bytes there, and a server that has only been
+  answering queries leaves `config.db-wal` empty and reads normally. Check an
+  `export` with `--config`, or stop the service first.
 
 ## `export`
 
@@ -92,17 +142,24 @@ See [back up and restore](../how-to/back-up-and-restore.md).
 ## `import FILE`
 
 Validates FILE and replaces the whole configuration with it in one transaction.
-Prints every validation problem on failure; a failed import leaves the database
-untouched. Refuses a database that already has content unless `--force` is
-given (`DatabaseNotEmpty`, exit 2). Creates the data directory at mode 0700 if
-it is missing.
+Prints every validation problem; a failed import leaves the database untouched.
+Refuses a database that already holds configuration unless `--force` is given
+(`DatabaseNotEmpty`, exit 2). The check measures what an operator set, not what
+the network did: client rows the DNS path materialised from traffic never
+trigger the refusal on their own. Creates the data directory at mode 0700 if it
+is missing.
+
+Client history survives the replacement. An address the database already knew
+keeps its first-seen and last-seen even when FILE names it; only an address it
+has never seen takes the import's clock. A client FILE leaves out is removed,
+history included.
 
 `FILE` is positional and may appear before or after the flags.
 
 | Flag | Meaning |
 | --- | --- |
 | `--data-dir DIR` | Data directory holding `config.db` (created if missing). |
-| `--force` | Replace a database that already has content. |
+| `--force` | Replace a database that already holds configuration. |
 
 ## `version`
 
@@ -120,31 +177,37 @@ the same. A usage error prints the same text to stderr and exits 64.
 | --- | --- |
 | 0 | Success. |
 | 1 | Runtime failure — I/O, database, out of memory. A partial diagnostic report caused by an allocation failure is a runtime failure, not a verdict on the configuration. |
-| 2 | A configuration problem the operator can fix, or a `check` that found one. Which faults qualify differs per subcommand; see below. |
+| 2 | A configuration problem the operator can fix, or a `check` that found one. |
 | 64 | Usage error — unknown command or flag, a flag without its value, a missing or extra argument. |
 
-Code 2 is not a single rule shared by every subcommand. `check` and `import`
-classify configuration faults; `run` maps only four errors to it and lets
-everything else out as a runtime failure.
+Code 2 means the same thing from every subcommand. `src/config/faults.zig`
+holds the one list of errors that mean "the configuration the operator supplied
+is wrong", and `run`, `check` and `import` all ask it, so a rejected file exits
+2 whichever command read it. The list is every error the validator raises, plus
+`ParseZon`, `ConfigTooLarge`, `NoUsableUpstreams` and `BadCertificate`. In
+practice that covers a seed file with a syntax error, one larger than 4 MiB, one
+with no `default` group (`MissingDefaultGroup`), one with no enabled upstream
+(`NoUpstreams`), a bad bind address, a bad rate limit, an unusable certificate,
+and `password` and `password_hash` set together.
 
-`check` exits 2 when validation reported at least one problem, when a
-certificate or key named by an enabled listener is unreadable, when a probed
-upstream failed, when the file is larger than 4 MiB, when the file has a ZON
-syntax error, and when there was nothing to check. Any other error escaping the
-run — a missing explicit `--config` file, an unreadable database — prints
-`check failed: <Error>` and exits 1.
+When `run` exits 2 it points at the diagnosis on stderr, whether the fault came
+from the seed file or from the database it loaded:
 
-`import` exits 2 when validation recorded at least one problem, and for
-`DatabaseNotEmpty`, `ConfigTooLarge`, `ParseZon` and `PasswordAndHashBothSet`.
-`OutOfMemory` is 1 even when problems were recorded, because the report is then
-incomplete. Every other error is 1.
+```
+run `nxdns check` to see the configuration in full
+```
 
-`run` exits 2 for `NoUsableUpstreams`, `BadBindAddress`, `BadRateLimit` and
-`BadCertificate`, and points the operator at `nxdns check` on stderr when it
-does. Nothing else is remapped. In particular, seeding the database from
-`--config` on first start happens before that classification applies to it: a
-seed file that is unparseable (`ParseZon`), larger than 4 MiB
-(`ConfigTooLarge`), or rejected by validation prints its diagnostics to stderr
-and exits 1. The same file given to `check` or `import` exits 2.
+`check` exits 2 for those faults and also when a probed upstream failed, when a
+named configuration file is missing or unreadable, when the database cannot be
+read or is not at this binary's schema version, and when there was nothing to
+check. Warnings never contribute.
+
+`import` exits 2 for those faults and for `DatabaseNotEmpty`. That last one is
+deliberately not a configuration fault — it reports the state of the database
+rather than the content of a file — and `import` decides it for itself; the
+answer to it is `--force`, not an edit.
+
+`OutOfMemory` is exit 1 even when problems were recorded, because the report is
+then incomplete. Every other error is 1.
 
 Where an exit code sends you next: [troubleshoot](../how-to/troubleshoot.md).

docs/reference/configuration.md

@@ -115,7 +115,7 @@ The DNS-over-HTTPS listener (server side, for clients on the LAN). See
 | `doh_server.enabled` | bool | false | — | — | gates the DoH listener (`src/server/doh_server.zig`) |
 | `doh_server.bind` | string | `"0.0.0.0"` | IP address | must parse as an IP address of either family | DoH listener bind |
 | `doh_server.port` | u16 | 443 | port | 1–65535 (0 is refused, enabled or not) | DoH listener port |
-| `doh_server.cert_path` | string | `"/etc/nxdns/cert.pem"` | path | non-empty when enabled | certificate loaded into the hot-reloading `CertStore`; readability is checked by `nxdns check`, not by the validator |
+| `doh_server.cert_path` | string | `"/etc/nxdns/cert.pem"` | path | non-empty when enabled | certificate loaded into the hot-reloading `CertStore`; the file itself is loaded and paired with the key by `nxdns check`, not by the validator |
 | `doh_server.key_path` | string | `"/etc/nxdns/key.pem"` | path | non-empty when enabled | private key for the same; `nxdns check` warns when it is readable beyond its owner |
 
 ### dot_server
@@ -169,7 +169,7 @@ the client tracker and the blocklist scheduler are throttled
 | Key | Type | Default | Unit | Validation | Consumed by |
 |---|---|---|---|---|---|
 | `blocklist_update.enabled` | bool | true | — | — | blocklist refresh scheduler (`src/filter/manager.zig`); when false the scheduler stops after the startup pass and only a manual refresh runs |
-| `blocklist_update.interval_hours` | u16 | 24 | hours | at least 1 | sleep between refresh passes and the per-source staleness test |
+| `blocklist_update.interval_hours` | u16 | 24 | hours | at least 1 | sleep between refresh passes and the per-source staleness test; each pass is preceded by the [orphan sweep](files-and-directories.md#the-orphan-sweep), which runs even when the pass itself is skipped for disk space |
 
 ## Collections
 
@@ -329,9 +329,35 @@ authentication](../how-to/set-up-admin-authentication.md).
 
 ## Validation errors
 
-`nxdns check` and `nxdns import` print one `path: message` line per problem and
-report every problem, not just the first. The error set is
-`validate.ValidateError` in `src/config/validate.zig`:
+`nxdns check`, `nxdns import` and the `nxdns run` that seeds a database from the
+file all print one `FAIL path: message` line per problem, and report every
+problem rather than the first. A finding that is legal but almost certainly
+unintended is prefixed `WARN` instead: it does not change the exit code, and it
+is printed by all three even when nothing failed, so an accepted configuration
+still says what is odd about it.
+
+A url in a diagnostic is redacted to its scheme, host and port. The userinfo,
+the path, the query and the fragment are dropped, and control characters are
+escaped. These lines reach the journal, and every one of those parts can carry a
+credential: a NextDNS DoH upstream is `https://dns.nextdns.io/abcd12`, where the
+path segment is the whole account identifier. The field path beside the message
+names the entry, so `blocklist_sources[1].url` still says which one to go and
+fix.
+
+One credential this cannot remove: a NextDNS **DoT** upstream is
+`tls://abcd12.dns.nextdns.io`, which carries the same identifier in the hostname.
+Stripping it would leave no host at all and no line worth reading. So if you use
+NextDNS over DoT, your profile id appears in the log.
+
+The host is kept because a hostname is not a secret in the general case — it is
+resolved publicly and offered as SNI on every connection — and dropping it would
+cost every operator a diagnostic to cover one vendor's choice. If that trade is
+wrong for you, it is yours to make rather than ours: NextDNS also publishes a DoH
+endpoint, `https://dns.nextdns.io/abcd12`, whose identifier sits in the path and
+is redacted in full. Configure that form instead and nothing identifying reaches
+the log.
+
+The error set is `validate.ValidateError` in `src/config/validate.zig`:
 
 | Error | Raised by |
 | --- | --- |
@@ -363,6 +389,13 @@ report every problem, not just the first. The error set is
 | `MissingLogPath` | `logging.output = .file` with an empty or relative `file_path` |
 | `PasswordAndHashBothSet` | both `web.password` and `web.password_hash` are set |
 
+Warnings are a separate set, outside `ValidateError` because they are not
+failures. There is one:
+
+| Warning | Raised by |
+| --- | --- |
+| `SourceInNoGroup` | a `blocklist_sources` entry that no `group_sources` link names; it downloads and blocks nothing until a group uses it |
+
 ## Minimal working example
 
 The smallest file that passes validation: a `default` group and one enabled

docs/reference/files-and-directories.md

@@ -16,19 +16,31 @@ never looks at the directory, and without one `check` falls back to the default
 configuration file, or prints "nothing to check" and exits 2 when neither source
 exists.
 
-No subcommand opens the directory read-only. `run`, `import`, `export` and a
-`check` that resolved to the database all go through the same
-`DataDir.openConfigDb`, which opens `config.db` read/write, chmods it to 0600,
-enables WAL — creating `config.db-wal` and `config.db-shm` — and then runs any
-pending schema migrations. So `nxdns export` and `nxdns check` write to the data
-directory, and on a database one schema version behind they migrate it. `export`
-additionally creates an empty `config.db` if the directory exists without one;
-`check` reaches the database branch only when `config.db` is already there.
+`run`, `import` and `export` go through `DataDir.openConfigDb`, which opens
+`config.db` read/write, chmods it to 0600, enables WAL — creating
+`config.db-wal` and `config.db-shm` — and then runs any pending schema
+migrations. So `nxdns export` writes to the data directory, and on a database
+one schema version behind it migrates it. `export` also creates an empty
+`config.db` if the directory exists without one.
+
+`nxdns check` is the exception: it does not use that path at all. It opens
+`config.db` immutable, which is `SQLITE_OPEN_READONLY` plus `immutable=1`, so
+SQLite refuses every statement that would write and builds no wal-index. No
+`config.db-wal` and no `config.db-shm` appear beside the file, the mode is left
+alone, and no migration runs — a database behind this binary's schema is
+reported as a failure naming `nxdns run` as the fix. A `check` against a data
+directory leaves it byte-identical, and `check` reaches the database branch only
+when `config.db` is already there.
+
+`immutable=1` ignores any `-wal` file, so it is refused rather than used when
+one holds bytes: the newest settings would be invisible and `check` would grade
+older ones from the main file. That is the "uncheckpointed changes" failure in
+[the CLI reference](cli.md#what-check-does-not-do).
 
 | Path | What it is | Mode |
 | --- | --- | --- |
 | `config.db` | The configuration database — the single source of truth, including `web.password_hash`. | 0600 |
-| `config.db-wal`, `config.db-shm` | SQLite write-ahead log and shared-memory index for `config.db`. Created when WAL is enabled, inheriting the main file's permissions. | 0600 |
+| `config.db-wal`, `config.db-shm` | SQLite write-ahead log and shared-memory index for `config.db`. Created by `run`, `import` and `export` when WAL is enabled, inheriting the main file's permissions. `check` creates neither. | 0600 |
 | `querylog.db` | The query log: every domain every client asked for. Expendable — if it is missing or unusable it is recreated empty. | 0600 |
 | `querylog.db-wal`, `querylog.db-shm` | WAL sidecars for `querylog.db`. | 0600 |
 | `querylog.db.corrupt-<unix-seconds>` | A `querylog.db` that could not be used, moved aside before an empty one was created in its place. Kept, never overwritten. | Whatever the renamed file had — no chmod reaches it |
@@ -38,11 +50,45 @@ additionally creates an empty `config.db` if the directory exists without one;
 | `blocklists/<id>.wild` | Wildcard entries for the same source. | 0600 |
 | `blocklists/<id>.raw.tmp`, `<id>.list.tmp`, `<id>.wild.tmp` | Transient refresh state: the downloaded body and the two compile outputs before they are published by rename. | 0600 |
 
-`<id>` is the `blocklist_sources` row id. A `.list` or `.wild` file whose id is
-no longer a `blocklist_sources` row is deleted by the orphan sweep; files of a
-live source are left alone whatever their state, and the sweep matches only the
-two published suffixes, so a `.tmp` file left by an interrupted refresh is not
-swept — the next refresh of that source overwrites it.
+`<id>` is the `blocklist_sources` row id.
+
+### The orphan sweep
+
+The sweep decides by id, not by suffix. It matches all five names above and
+deletes those whose `<id>` is no longer a `blocklist_sources` row, so the
+compiled `.list` and `.wild` of a removed source go, and so do a `.raw.tmp`,
+`.list.tmp` or `.wild.tmp` left behind by a refresh that was killed before it
+could clean up. Files belonging to a source that still has a row are never
+touched, whatever state they are in: the sweep holds the same lock every refresh
+takes, so it never reads the directory while a refresh is part-way through.
+
+It runs at three moments:
+
+- at startup, before the first refresh pass — this is what collects what a
+  killed process left behind, and the compiled files of a source deleted while
+  the server was down;
+- before each scheduled update pass, ahead of the disk-space gate: the sweep
+  only unlinks, so it is the one step here that can give a critically full disk
+  room back, and gating it would keep the residue that helped fill the disk;
+- immediately after `DELETE /api/blocklists/{id}`, which is when an orphan is
+  actually created in normal operation. Without it a deleted list would keep its
+  megabytes until the next scheduled pass.
+
+With `blocklist_update.enabled = false` there are no scheduled passes, so only
+the first and the last of those three happen.
+
+Each deletion is logged:
+
+```
+info(blocklist_manager): pruned orphaned blocklist file 9999.list
+```
+
+A failed sweep is a warning, not an outage — leftover bytes do not justify
+losing the refresh pass behind them, let alone the server.
+
+The temporaries of a source that still exists are cleaned by the refresh that
+owns them rather than by the sweep: each refresh deletes its own `.raw.tmp`,
+`.list.tmp` and `.wild.tmp` as it finishes, successfully or not.
 
 A `querylog.db` is moved aside when it is missing nothing but usability:
 SQLite reports it corrupt or not a database, `PRAGMA quick_check` does not
@@ -111,7 +157,9 @@ no file; under systemd the journal captures it.
 
 `doh_server.cert_path` / `key_path` and `dot_server.cert_path` / `key_path`,
 conventionally under `/etc/nxdns`. nxdns reads them, never writes or creates
-them. Both must be readable by the user nxdns runs as; the key should be
-readable by its owner only, which `nxdns check` warns about when it is not. A
+them. Both must be readable by the user nxdns runs as, and the key must belong
+to the certificate: `nxdns check` loads the pair and fails when it does not. The
+key should also be readable by its owner only, which `check` warns about when it
+is not. A
 watcher polls both files and swaps a renewed pair in without a restart. See
 [enable DoH and DoT](../how-to/enable-doh-and-dot.md).

docs/reference/performance.md

@@ -12,8 +12,8 @@ see [measure performance](../how-to/measure-performance.md).
 | Sustained ≥ 100 qps on Raspberry Pi 5 | End-to-end against the real binary on the Pi; not a harness number |
 | Blocklist lookup p95 < 1 ms | `bench filter`: `matcher.normalize` + `Snapshot.evaluate` per op |
 | Cached response p95 < 5 ms | `bench cache`: `buildKey` + `DnsCache.get` + `packet.setId` per op |
-| Memory with ~1M blocked domains < 100 MB | `bench filter`: VmRSS with the 1M-domain snapshot loaded |
-| Stripped static binary < 10 MB per arch (< 15 MB with the embedded frontend) | CI size assert on the `cross` artifacts |
+| Memory with ~1M blocked domains < 100 MiB | `bench filter`: VmRSS with the 1M-domain snapshot loaded |
+| Stripped static binary < 10 MiB per arch (< 15 MiB with the embedded frontend) | CI size assert on the `cross` artifacts |
 
 The harness is `tools/bench.zig`. It measures the three targets that are
 measurable in process; the qps target is end to end and the binary-size target
@@ -61,7 +61,7 @@ VmRSS is lower because the filter suite's snapshot has been freed by then.
 | --- | --- |
 | Blocklist lookup p95 < 1 ms | to be measured on hardware |
 | Cached response p95 < 5 ms | to be measured on hardware |
-| Memory with ~1M blocked domains < 100 MB | to be measured on hardware |
+| Memory with ~1M blocked domains < 100 MiB | to be measured on hardware |
 | Sustained ≥ 100 qps | to be measured on hardware, end to end |
 
 The qps target belongs to the real binary rather than the harness: it means

docs/tutorial/first-run.md

@@ -71,9 +71,9 @@ an install decision, not a first-run decision. A real install is covered in
 The `default` group and one enabled upstream are the two things nxdns will not
 start without. Every client that nxdns has never seen is assigned to `default`,
 and with no usable upstream there is nowhere to send a query it cannot answer
-itself, so a configuration missing either one is rejected. Seeding happens
-inside `run`, so `run` prints the problem and exits 1; `nxdns check` and `nxdns
-import` reject the same file with exit code 2.
+itself, so a configuration missing either one is rejected. Whichever command
+reads the file says the same thing and stops the same way: `run`, `nxdns check`
+and `nxdns import` all print the problem and exit 2.
 
 ## 4. Check the configuration before starting
 
@@ -83,13 +83,14 @@ zig-out/bin/nxdns check --data-dir ~/nxdns-tutorial/data --config ~/nxdns-tutori
 
 ```
 checking configuration file /home/you/nxdns-tutorial/config.zon
-OK https://cloudflare-dns.com/dns-query
+OK upstreams[0] https://cloudflare-dns.com
 OK: no problems found
 ```
 
 `check` parses the file, validates it, and contacts each upstream to confirm it
-answers. It exits 0 when there is nothing to fix and 2 when there is. It does
-not start any listener, so you can run it as often as you like.
+answers. It exits 2 when it found something that has to be fixed and 0
+otherwise. It writes nothing and starts no listener, so you can run it as often
+as you like.
 
 ## 5. Start the server
 
@@ -158,6 +159,13 @@ separate decision, which is what lets one group get a strict list and another
 get none. A source that is attached to no group is downloaded and then filters
 nothing.
 
+You are in that state right now, between the previous step and this one, and it
+is legal rather than wrong — creating a source and attaching it afterwards is
+the normal order, which is why the API accepted it without complaint. On a
+stopped server `nxdns check` names such a source in a `WARN` line and still
+exits 0, so a list that silently blocks nothing is something you can find out
+about later. Attaching it now is what makes it take effect.
+
 Ask which groups exist:
 
 ```sh

specs/milestone-13.md

@@ -136,9 +136,56 @@ Five review rounds: 15 findings, then 8, then 1, then 1, then clean. One round-1
 finding was rejected on evidence (`troubleshoot.md` already stated the verified
 exit codes; the contradiction was in `reference/cli.md`).
 
-## Discrepancies found, not fixed
+## Fix wave (supersedes ruling 9)
 
-Docs-only per ruling 9. These are source-side and belong to a later milestone:
+Ruling 9 kept the documentation wave docs-only so that writing pages could not
+churn behaviour underneath itself. That job is done and committed at 16c9de2.
+The nine discrepancies the wave found are now fixed inside milestone 13, on the
+user's direction; milestone 14 stays packaging and publishing.
+
+Binding rulings for the fix wave:
+
+F-a. **One definition of "the operator's configuration is wrong."** The
+divergence in D1 exists because `app.isConfigFault` and `cli.failureExitCode`
+each carry their own list. Neither list is the fix. Add `src/config/faults.zig`
+with exactly:
+
+```zig
+pub fn isConfigFault(err: anyerror) bool
+```
+
+covering the seed and validation errors (`ParseZon`, `ConfigTooLarge`,
+`MissingDefaultGroup`, `NoUpstreams`, `NoUsableUpstreams`, `BadBindAddress`,
+`BadRateLimit`, `BadCertificate`, `PasswordAndHashBothSet`, and every
+`validate.ValidateError`). `app.zig` and `cli.zig` both call it and keep no
+private list. `run`, `check` and `import` then agree: a rejected configuration
+file is exit 2 from every subcommand.
+
+F-b. **Diagnostics carry severity.** `validate.Diagnostics` gains a severity per
+item — `.fail` or `.warn`. `writeAll` prints `FAIL `/`WARN ` accordingly.
+Counting splits: failures set exit 2, warnings never change an exit code. This
+is the API D2 and D5 both need, so it is pinned here rather than invented twice.
+
+F-c. **`check` never writes.** It opens `config.db` read-only and does not
+migrate. A database behind the current schema is reported, not upgraded. If a
+read-only open is impossible for a database needing WAL recovery, report that as
+a failure naming `nxdns run` as the fix — do not silently fall back to a
+writable open.
+
+F-d. **`check` proves what it claims.** It parses the certificate and the key and
+verifies they pair, through the same code the server uses, so a green `check`
+cannot be followed by `run` exiting 2 on `BadCertificate`.
+
+F-e. **Documentation follows behaviour in the same wave.** Every page that
+documents the old behaviour is corrected: exit codes in `reference/cli.md`,
+`how-to/troubleshoot.md`, `how-to/install-with-docker.md`, `tutorial/first-run.md`;
+the certificate gap in `how-to/troubleshoot.md` and `how-to/enable-doh-and-dot.md`;
+the "check writes" note in `reference/files-and-directories.md`. The three drift
+guards stay green. Ruling 3 still binds: a changed command is re-run here.
+
+F-f. **Every fix ships with a test that fails without it.**
+
+## The nine discrepancies
 
 1. `nxdns run` exits 1 for seed-file errors (`ParseZon`, `MissingDefaultGroup`,
    `NoUpstreams`) while `check` and `import` exit 2 for the same file, because
@@ -161,6 +208,373 @@ Docs-only per ruling 9. These are source-side and belong to a later milestone:
    one query ignores a seed file placed afterwards.
 9. PLAN §18 and the docs say 10 MB / 15 MB; CI asserts 10 MiB / 15 MiB.
 
+## Fix wave delivered
+
+All nine are closed. Each shipped with a test its author watched fail with the
+implementation reverted — ruling F-f was enforced by demanding the observed
+failure output, not an assertion that a test would fail.
+
+D1 is one `src/config/faults.zig` deriving the fault set by comptime reflection
+over `validate.ValidateError`, with no exclusion list; `run`, `check` and
+`import` all exit 2 on the same rejected file, for `MissingDefaultGroup` and for
+`ParseZon`. D2 and D5 rest on `validate.Diagnostics` gaining `.fail`/`.warn`:
+`check` prints `OK: no failures found, 1 warning` and exits 0. D3 proves the
+certificate pair through `cert_store.CertStore.init`, the same code the
+listeners use, so a green `check` cannot be followed by `run` exiting 2 on
+`BadCertificate`. D4 fixes the missing-file exit at the root read. D6 opens
+`config.db` with `OpenMode.immutable` and no migrate; a stale `-wal` is reported
+naming `nxdns run`, never read past. D7 sweeps all five suffixes. D8 counts
+`hand_edited = 1` only. D9 is MiB everywhere CI asserts MiB.
+
+Four defects the wave found that were not among the nine:
+
+1. **`pruneOrphans` had no production caller.** D7's widened matching was
+   unreachable at runtime. `Manager.sweepOrphans` now runs at startup, before
+   each interval pass, and on `DELETE /api/blocklists/:id`.
+2. **An import destroyed a device's observed timestamps.** `first_seen` and
+   `last_seen` are runtime state, not configuration, and the model carries no
+   field for either — so a configured row was inserted with the import clock in
+   both columns. They now follow the address: a device the database already knew
+   keeps them, only an unseen address takes the import's clock, and a client the
+   file omits is removed with its history. Verified live, not only in tests.
+3. **A blocklist or upstream url reached the log whole.** A signed url or an
+   `?apikey=` query persisted in journald. `src/safe_url.zig` exports `redact`,
+   which keeps scheme, host and port and drops userinfo, path, query and
+   fragment. It scans rather than parses, deliberately: `error.BadUrl` is one of the
+   failures these very lines report, so the inputs a parser refuses are exactly
+   the ones that must still redact. Twenty-two call sites across `manager.zig`,
+   `validate.zig`, `cli.zig` and `app.zig`.
+4. **`src/main.zig` built both runner writers in positional mode.** With stderr
+   redirected to a regular file, runner output pwrote over what `std.log` had
+   already written at offset 0, and `2>>` was silently broken because pwrite
+   ignores `O_APPEND`. Both writers now use `writerStreaming`.
+
+Two review findings were answered against the reviewer rather than by it, both
+with evidence rather than argument:
+
+- **A failed stderr write must not stop the server.** The reviewer wanted
+  `seedFromFile` to propagate an output failure as a runtime failure. Implementing
+  that proposal and running the suite showed it replacing
+  `error.MissingDefaultGroup` with `error.WriteFailed` — a broken stderr would
+  hide why the seed file was refused. The discards stay, and the reasoning now
+  sits above them, with the derived half of the claim labelled as derived.
+- **The log-injection hole was already closed for `std.log`.** `logging.zig:342`
+  escapes control bytes in every log message, so the forged-line scenario the
+  reviewer described could not happen through that path. The live gap was
+  `cli.zig`'s stdout. Escaping stays in `safe_url` as well as the sink, because
+  `validate.Diagnostics` builds `Problem.message` as an allocated string that
+  `web/handlers/mutations.zig` returns as a 400 body — a channel no log sink can
+  escape.
+
+Redaction ended stricter than it started. `redact` prints scheme, host and port
+only: a NextDNS DoH upstream is `https://dns.nextdns.io/abcd12`, where the path
+segment is the whole account identifier, so keeping the path kept the credential.
+The rule has no exemption for `nxdns check`'s stdout, which is the output an
+operator pastes into a bug report. Log lines identify a source by row id and name
+instead, and the two duplicate diagnostics now name the other entry
+(`duplicate of upstreams[0]`) rather than quoting a url that no longer shows why
+the two collide.
+
+One credential redaction cannot remove, pinned as a test and documented rather
+than hidden: a NextDNS **DoT** upstream is `tls://abcd12.dns.nextdns.io`, which
+carries the same identifier in the hostname. Removing it would leave no host and
+no actionable line. A hostname is resolved publicly and offered as SNI in any
+case, so it is not private the way a query string is, and dropping every host
+would cost every operator a diagnostic to cover one vendor's choice. The
+reference page names the mitigation the operator controls instead: NextDNS also
+publishes a DoH endpoint whose identifier sits in the path and is redacted whole.
+
+Four rounds of review hardened `redact` and each one found what the last missed.
+Three of the four were the same root cause, which is worth naming because the
+first two fixes treated it as bad luck: **the scanner tried to identify a host
+inside text that is not a url, and guessed.** It guessed the query was safe to cut
+before the userinfo, leaving `user:pa55` as the host. It guessed a `\` was not a
+separator. And it guessed which side of an `@` was the host when a late delimiter
+made both readings available.
+
+That third one was recorded here, in an earlier revision of this section, as an
+accepted cost: an `@` in a backslash path prints "a misleading host, never a
+credential". **That claim was false and is retracted.** Running the shipped code
+disproved it:
+
+```
+https://lists.example?token=prefix@hunter2  ->  https://hunter2
+https://lists.example#f@hunter2             ->  https://hunter2
+https:\\lists.example\p@hunter2             ->  https://hunter2
+```
+
+A query string is the most likely place in a url for a token, so the text the
+scanner promoted to "host" was the secret itself. The note claiming otherwise is
+why three subsequent rounds passed over it.
+
+The fix is a rule rather than a fourth special case: **the scan no longer guesses,
+it declines.** The `/` cut now runs first and unconditionally, which is safe for
+the reason the old ordering missed — everything after the `/` is dropped anyway,
+so an `@` there never needed to be userinfo to stay out of the log. Where two
+readings genuinely survive, the authority is omitted whole and `format` prints
+`(ambiguous authority omitted)`, which is prose rather than a placeholder host so
+an operator reads it as a statement about the line. `SafeUrl.authority` is
+therefore `?[]const u8`: empty and `null` are different answers, one saying the
+url names no authority and the other saying it names one that cannot be resolved.
+The same rule caught a case nobody had raised — `https:a@hunter2`, where RFC 3986
+reads `hunter2` as a path segment and WHATWG reads it as the host. The
+disagreement between two parsers is itself the evidence of ambiguity.
+
+A redesign was considered and rejected: parse with `std.Uri.parse` first and print
+nothing but the scheme when the parse fails. It reaches the same "do not guess"
+place, but it also discards the host for every url malformed in a harmless way,
+and `redact` exists to be callable from the `error.BadUrl` paths that report
+exactly those. The ordering fix gets the property without the cost.
+
+Escaping is owned by the type that introduces the delimiter: `quoteText` writes
+its own quotes and escapes `'` inside them, so a caller cannot reopen the hole by
+adding quotes of its own. Escaping inside a plain helper would have left the
+defect one caller away, which on a third review round is not a fix.
+
+`SafeUrl` was then found to break that same rule from the other side. Its `format`
+hardcoded the `none` delimiter, so it never escaped `'` — while eight call sites
+wrapped it in `'{f}'` of their own. `https://ho'st/x` redacts to `https://ho'st`,
+which inside a caller's quotes reads as `'ho'` followed by loose text. Two
+independent findings converged on it: this review, and the `/metrics` work, which
+hit the same shape with `"` instead of `'`.
+
+`redactQuoted` closes it, and the choice between the two is a stated rule rather
+than per-call-site judgement: **use the quoted form whenever anything follows the
+url on the line**, because a redacted authority can still hold a space, a `:` and a
+`'`, and unquoted it can impersonate whatever comes next — `upstream {f} failed:
+{t}` with an authority of `ok failed: Timeout` reports a failure that did not
+happen. Bare `redact` is for the two cases where that cannot arise: the url ends
+the line (`cli.zig`'s `OK upstreams[N]`, `context.zig`'s missing-id line), or the
+caller owns the escaping for a delimiter of its own (`metrics.zig`). Ten sites
+moved to the quoted form. The output is byte-identical for any url without a `'`,
+so no documented output changed.
+
+One property is asserted rather than assumed, because it is what makes a `\` in the
+output always this file's and never the operator's: a `\` ends an authority, so
+unlike a source name it can never reach the value to be doubled.
+
+Round five found two more, both in the same scanner, which is now five rounds and
+five findings. Both were confirmed by running the code before being fixed.
+
+The first is the ambiguity rule applied to only half its cases. `https:a@hunter2`
+was withheld, but `https:hunter2` — the same opaque path with no `@` in it —
+printed whole, because the check sat *after* the `@` lookup and a url with no `@`
+returned before reaching it. The reading is ambiguous either way; the `@` was never
+what made it so. The check now runs before the `@` lookup.
+
+Moving it exposed the reason it had been placed there: `localhost` satisfies the
+scheme production, so a rule keyed on the production alone withholds
+`localhost:8080/x`, which is the shape an operator on a LAN is most likely to
+write. A port is not an opaque path, so the digits after the colon are what
+separate the two. This is checked against the trimmed authority, not the raw one,
+or `localhost:8080?x` would fail the digit test on the query.
+
+The second: a network-path reference (`//lists.example/hosts.txt`) redacted to the
+empty string, because the unconditional `/` cut lands at byte zero. Not a leak —
+nothing was printed — but the line named no source at all, and the authority in it
+is not in doubt. A leading run of `/` is now consumed the way a scheme delimiter's
+is, so the userinfo in `//user:pa55@lists.example/x` is dropped rather than the
+whole authority withheld. An ambiguous one such as `//a?b@c` is still withheld: the
+branch settles where the authority starts, not that every reading of it is
+resolved.
+
+Round six found three more in the same function, all the same class again: a
+separator run that was read as an authority delimiter when it does not settle one.
+
+`https:/hunter2` and `https:///hunter2` printed the path segment as the host. The
+scan accepted a delimiter of *any* number of separators, and the comment recording
+why was accurate when it was written and stale by the time it was read: the
+tolerance existed so `https:/user:pass@host/list` would find an authority instead
+of printing its userinfo. That reason expired when the `/` cut moved ahead of the
+userinfo lookup in round four — the authority of a url with no `://` now ends at
+its first `/`, so it holds no userinfo to print. Only a run of exactly two
+introduces an authority; one leaves an absolute path and three or more is an empty
+authority to RFC 3986 and a host to WHATWG. The four inputs that motivated the old
+tolerance are still safe, now by being withheld rather than resolved, and that is
+asserted where the old behaviour used to be.
+
+`///lists.example/x` had the same defect in the network-path branch added one
+round earlier, which consumed the whole run. Exactly two there too.
+
+The third retracts something this section claimed one round ago. `isPort` was
+introduced so a schemeless `localhost:8080` would keep resolving, on the reasoning
+that a port is not an opaque path. It is not that simple: `https:123456` is an
+opaque path whose digits are as much a token as any other text, and the exception
+printed it whole. The reviewer proposed excluding known schemes from the
+exception; the exception is gone instead, because a list of known schemes is the
+kind of thing that goes stale silently and this file already has one rule that
+covers it. `localhost:8080/x` is now withheld, which costs nothing an operator
+needs: a url reaching this without a scheme is one the validator is rejecting, and
+the field path beside it names which. An `IP:port` is unaffected — a leading digit
+fails the scheme production, so `10.0.0.2:8080` and `[::1]:853` still resolve.
+
+Round seven took the same rule one step further. "Exactly two separators" still
+accepted `\\`, `/\` and `\/`, and WHATWG converts a `\` to a `/` only for a
+*special* scheme — so `https:\\hunter2` has the same split reading as the cases
+above, and `tls:\\hunter2` has no reading at all under which `hunter2` is a host,
+`tls` not being special. RFC 3986 gives `\` no meaning anywhere. Only `//` opens
+an authority now. A `\` still ends one, and still anchors the scheme scan so a
+backslash-pasted url is reported by its scheme, but it opens nothing.
+
+Round seven also caught a stale comment, which is the second time in two rounds
+that a comment outlived the reason it recorded. Both said something true when
+written and false when read, and both were load-bearing — the first is why the
+any-length tolerance survived three rounds after its justification expired.
+
+Seven rounds, nine findings, one function. Every one was a place where the scan
+committed to a reading the text did not support. What finally holds is not a
+sharper scan but a smaller claim: the authority is printed only where exactly one
+reading survives, and withheld everywhere else. Concretely, an authority is read
+only after a literal `//`, whether a scheme introduced it or not.
+
+The cost is paid by malformed input alone, and it is paid in diagnostic detail
+rather than in safety: a backslash-pasted or schemeless url now reports its scheme
+and no host. Every such url is one the validator is already rejecting, and the
+field path or row id beside it names which one. A sweep of eighteen shapes
+carrying `hunter2`, `abcd12`, `s3cr3t`, `token`, `pa55` and `123456` finds none of
+them in any output.
+
+Round eight returned no critical and no important finding, and two minor ones that
+cut in opposite directions. A run of three or more slashes was reported as an
+authority the url does not have, when it is contested rather than absent — RFC
+3986 reads an empty authority, WHATWG resolving against a special-scheme base
+reads a host. `SafeUrl` exists to keep those two answers apart, so it is now
+withheld rather than reported as empty.
+
+The other corrected a claim this file made about the very thing it was withholding.
+`localhost:8080` is **not** contested: `localhost` is not one of WHATWG's six
+special schemes, so both standards read a scheme and an opaque path, and the
+host-and-port an operator meant is a reading no parser offers. It is still
+withheld — the text after the colon is a path segment under every reading — but the
+marker calls it an authority that could not be resolved when the honest answer is
+that there is none.
+
+That wording defect was first recorded rather than fixed, on the grounds that
+telling the two apart needs a scheme list and a scheme list is what the previous
+round had just removed. Round nine rejected that and was right to: the list round
+six removed described *what nxdns supports*, so it went stale whenever a transport
+was added. WHATWG's special schemes are a closed set fixed by the URL Standard —
+`ftp`, `file`, `http`, `https`, `ws`, `wss` — which never described nxdns and
+cannot go stale with it. Conflating the two was the error, and knowingly shipping
+a diagnostic that contradicts its own type's contract is the tech debt this
+project does not take on.
+
+So `redact` now distinguishes them. Only a special scheme can disagree with RFC
+3986 about text no `//` introduced, so `https:hunter2` is withheld as contested
+and `localhost:8080`, `mailto:ops@example.com` and `tls:\\host` are withheld as
+naming no authority at all. Round nine's other finding was the same rule missing
+from the schemeless branch: `\\lists.example\path` and `/\lists.example/path` were
+reported as absent when they are contested. A run is settled when it is one
+separator, or when the reading that looks for a host finds none —
+`\\?\C:\lists\hosts.txt` ends its authority at the `?` under both — and contested
+otherwise.
+
+Round ten returned nothing critical and nothing important, four minor findings and
+a nit. Two were taken: the nit, which was a comment claiming
+`https:\\dns.nextdns.io\abcd12` "names a host" when the file's own reasoning is
+that it is contested; and the rendering question — `tls:\\host` prints `tls://`
+although the input held no `//`. That one is now contractual rather than
+accidental. `format` renders `scheme://authority` canonically and does not quote
+the input's syntax; nothing about which bytes separated the scheme survives
+redaction, and nothing should, because the input is not meant to be reconstructible
+from the output.
+
+**Three findings were declined, and the reason is checked rather than asserted.**
+`https:/user@/x`, `\path@hunter2` and `file:secret` are classified as contested
+when both readings in fact find no host. Running them shows why that is tolerable:
+each prints `(ambiguous authority omitted)`, so each says *less* than it could and
+none prints the path segment. The mechanisms are named in the file — emptiness
+tested before userinfo removal, a leading run of one reaching the late-delimiter
+rule, and `file` having its own WHATWG parsing states that `isSpecialScheme` does
+not model. Fixing them means modelling more of two standards for inputs no
+accepted configuration can hold: every url this program takes carries `http`,
+`https`, `tls`, `udp` or `tcp` and a `//`. All three are pinned by a test, because
+the failure that would matter is the opposite one — if any of them ever starts
+naming a host, that test fails.
+
+That is where the review loop was stopped, on a judgement rather than on an empty
+round. Rounds eight, nine and ten each returned nothing critical and nothing
+important, and the findings had moved from "this prints a secret" to "this claims
+more than it knows about a url no operator can configure". Round ten opened a new
+sub-class of its own — WHATWG's `file:` state machine — which is the signal that
+the remaining work is unbounded and no longer about safety.
+
+Ten rounds, sixteen findings, one function. A sweep of twenty-five shapes carrying
+`hunter2`, `abcd12`, `s3cr3t`, `token`, `pa55`, `123456`, `hosts.txt` and
+`example.com` finds none of them in any output.
+
+The last gap was found by sweeping every log site in the tree rather than
+trusting the review's file list: `src/upstream/dot_client.zig` printed the
+upstream url whole at four sites and `pool.zig` at a fifth. Three review rounds
+and four agents had redacted urls across `manager.zig`, `validate.zig`, `cli.zig`
+and `app.zig` without anyone asking which other modules logged the same values.
+
+Deliberately not changed: `POST /api/blocklists` still ignores the
+`SourceInNoGroup` warning. The web flow creates a source before any group link
+can exist, so the warning is structural at that moment and a 400 would be wrong.
+The three CLI paths carry it instead.
+
+Claims in the wave that no test backs, recorded rather than buried: the two
+`app.zig` `log.warn` redactions, because a `std.log` line is not observable from
+a unit test under the default runner; `cli.zig`'s `not a usable DoH url`
+redaction, argued unreachable by any credential-carrying url because
+`Endpoint.parse` rejects `@?#` in the authority first; the two probe lines that
+need a reachable upstream; and the WAL guard's residual window, which is argued
+rather than observed.
+
+## Round four
+
+The `/metrics` family was the last place a credential still reached an
+unauthenticated reader. `GET /metrics` is `.auth = .open` in `src/web/routes.zig`
+and `web.bind` defaults to `0.0.0.0`, so an unauthenticated `curl` printed
+`nxdns_upstream_up{url="https://dns.nextdns.io/abcd12"} 1` with HTTP 200. Every
+other redaction in this wave was on a path that already required a session or a
+shell on the host; this one was not.
+
+Redacting it needed a second escaping layer rather than a call to `redact`.
+`redact` cuts the authority at `/@?#\` and escapes every control byte, so it can
+emit neither a raw newline nor a lone backslash — but `"` is in none of those cut
+sets, so `https://ho"st/x` arrives at the label as `https://ho"st` and closes the
+label value, letting the rest of the string write label pairs of its own.
+`writeLabelValue` therefore runs *over* the redacted text. The ordering also
+matters in the other direction: doubling the backslash turns the two characters
+`redact` writes for a control byte into an unambiguous `\\n`, so a parser reads a
+backslash rather than a newline.
+
+The redaction then introduced a defect of its own, caught before it shipped: two
+upstreams on one host redact to one label set, so `nxdns_upstream_up` rendered
+twice with identical labels in a single exposition. Two NextDNS profiles is a
+realistic configuration. The fix labels each sample with its array position in
+`renderUpstreams`, so uniqueness holds by construction and a hand-built `Sample`
+cannot forge a collision. The index is positional and therefore stable only while
+the pool order is — `pool.Snapshot` carries no row id, and threading one through
+the pool and the repository was judged out of scope for a review round. The
+caveat is in the file. It only bites when two upstreams share an origin; for
+distinct origins the url carries the identity and a reorder is harmless.
+
+Claim scoping on that defect, since the three statements are not one: the
+duplicate exposition text **was** reproduced, out of the same `render` function
+`/metrics` calls. It was **not** reproduced over HTTP from a pre-fix server — every
+live run had the index in place. And "a scrape carrying it is rejected" is
+**reasoned, not observed**; nothing was fed to a Prometheus. Whether a scraper
+rejects the pair or keeps the last sample, the down upstream is lost, which is the
+part the fix rests on.
+
+`src/storage/repositories/context.zig` printed a blocklist url whole and a group
+name raw on its two `NotFound` paths. Both are invariant-failure paths that fire
+rarely, which is why they outlived three rounds of sweeping.
+
+One test file was never running. `src/config/faults.zig` was absent from
+`src/tests.zig`, and Zig collects tests only from the root module's explicit
+import list — a transitively imported file contributes none, even when its values
+are used, which was confirmed against a scratch project rather than assumed. The
+five reflection guards in that file are what catch a future `ValidateError`
+variant being added and left unclassified; they had never executed. Its runtime
+behaviour was covered — `cli.zig` has a test that classifies through it — so the
+gap was in the regression guard, not in what shipped. A sweep of the tree found no
+other unlisted file with tests; `dns/dns.zig` is a re-export barrel with none, and
+its own comment documents this rule.
+
 ## Acceptance (milestone complete)
 
 - [x] The five old `docs/*.md` files are gone; the four directories and the index
@@ -168,9 +582,9 @@ Docs-only per ruling 9. These are source-side and belong to a later milestone:
 - [x] Every command in `tutorial/` and `how-to/` was executed on this host, or is
       marked in-page as unverified with a reason.
 - [x] `zig build test` passes with the three drift guards pointing at the new
-      reference pages: 16/16 steps, 1175 pass, 113 skip (integration-gated), 0 failed.
-- [x] `zig build test -Dintegration` passes: 1284 pass, 4 skip (the live-network
-      TLS tests excluded by milestone-1 design), 0 failed. No src/ behavior change.
+      reference pages: 16/16 steps, 1254 pass, 115 skip (integration-gated), 0 failed.
+- [x] `zig build test -Dintegration` passes: 16/16 steps, 1365 pass, 4 skip (the
+      live-network TLS tests excluded by milestone-1 design), 0 failed.
 - [x] No page mixes modes: no procedure in `reference/`, no field table in any
       `how-to/` or `tutorial/` page.
 
@@ -179,5 +593,6 @@ Docs-only per ruling 9. These are source-side and belong to a later milestone:
 - No new documentation subjects (no metrics-families reference, no security-model
   page) — this milestone moves and splits existing content plus the tutorial.
 - No redirect stubs or `docs/legacy/`.
-- No source fixes; report discrepancies instead.
+- No source fixes; report discrepancies instead. (Held for the documentation
+  wave. Superseded by the fix wave above, which closes the nine it reported.)
 - No doc generator, no site builder, no MkDocs.

src/app.zig

@@ -44,6 +44,7 @@ const doh_client = @import("upstream/doh_client.zig");
 const doh_server = @import("server/doh_server.zig");
 const dot_client = @import("upstream/dot_client.zig");
 const dot_server = @import("server/dot_server.zig");
+const faults = @import("config/faults.zig");
 const fetcher = @import("filter/fetcher.zig");
 const forward_zones = @import("local/forward_zones.zig");
 const handler = @import("server/handler.zig");
@@ -60,6 +61,7 @@ const pool_mod = @import("upstream/pool.zig");
 const query_sink = @import("server/query_sink.zig");
 const rate_limiter = @import("server/rate_limiter.zig");
 const retention_mod = @import("storage/retention.zig");
+const safe_url = @import("safe_url.zig");
 const shutdown = @import("server/shutdown.zig");
 const sse = @import("web/sse.zig");
 const static = @import("web/static.zig");
@@ -87,24 +89,14 @@ const download_budget_s = 300;
 const doh_request_buf_len = 1024;
 const doh_transfer_buf_len = 4096;
 
-/// A configuration fault the operator can fix, as opposed to a runtime one.
-/// These are the only failures this file raises itself; everything else comes
-/// out of a collaborator.
-const ConfigError = error{
-    NoUsableUpstreams,
-    BadBindAddress,
-    BadRateLimit,
-    BadCertificate,
-};
-
 pub fn run(runner: cli.Runner, args: cli.RunArgs) u8 {
     const code = serve(runner, args) catch |err| code: {
         runner.err.print("nxdns run failed: {s}\n", .{@errorName(err)}) catch {};
-        if (isConfigFault(err)) {
+        const mapped = failureExitCode(err);
+        if (mapped == cli.exit_check) {
             runner.err.writeAll("run `nxdns check` to see the configuration in full\n") catch {};
-            break :code cli.exit_check;
         }
-        break :code cli.exit_runtime;
+        break :code mapped;
     };
 
     // Output the operator never received is not output, so a failed flush
@@ -114,15 +106,75 @@ pub fn run(runner: cli.Runner, args: cli.RunArgs) u8 {
     return code;
 }
 
-fn isConfigFault(err: anyerror) bool {
-    return switch (err) {
-        error.NoUsableUpstreams,
-        error.BadBindAddress,
-        error.BadRateLimit,
-        error.BadCertificate,
-        => true,
-        else => false,
-    };
+/// The one classification, from `config/faults.zig`. This file keeps no list of
+/// its own: `run` exiting 1 on a seed file `check` and `import` exit 2 on was
+/// exactly the cost of the second list that used to be here.
+fn failureExitCode(err: anyerror) u8 {
+    return if (faults.isConfigFault(err)) cli.exit_check else cli.exit_runtime;
+}
+
+/// First run only: the file seeds an empty database and is ignored forever
+/// after. Its diagnostics are the operator's one chance to see what the file
+/// said, so they are printed the way `check` and `import` print them.
+///
+/// Printed on the way out either way. A seed file can be accepted and still
+/// carry warnings — a blocklist source in no group is downloaded and compiled
+/// into nothing — and a warning that only appears when the start fails is a
+/// warning nobody ever reads: the start it describes is the one that worked.
+/// `check` reported it and `run` did not, which left the same file graded two
+/// ways.
+///
+/// The runner's error writer, not `std.log`: this runs before
+/// `logging.install`, and one rendering of a diagnostic across `run`, `check`
+/// and `import` is the point of `Diagnostics.writeAll`.
+///
+/// Flushed here rather than left to `run`'s exit flush. That writer is buffered
+/// (`main` gives it 4 KiB) and `serve` does not return for as long as the
+/// service runs, so a line left in the buffer reaches the operator when the
+/// process stops — days after the start it describes. A failure path flushes
+/// anyway because it returns immediately; the successful start is the one that
+/// needs this.
+fn seedFromFile(
+    r: cli.Runner,
+    config_db: *db.Db,
+    dir: std.Io.Dir,
+    config_path: []const u8,
+) bootstrap.Error!bootstrap.Outcome {
+    var diags: validate.Diagnostics = .init(r.gpa);
+    defer diags.deinit();
+
+    const result = bootstrap.bootstrap(r.io, r.gpa, config_db, dir, config_path, &diags);
+
+    // Neither discard is an oversight, and the two answer different questions.
+    //
+    // On a rejected seed file, `result` is returned untouched: the operator gets
+    // the reason the start failed, never a writer error standing in front of it.
+    // A broken stderr is not why the configuration was refused.
+    //
+    // On a seed that worked, a failure here does not stop the start. The trade
+    // is one lost warning line against a household with no name resolution, and
+    // `run` before this point is the only stretch of this program where an
+    // output failure could take DNS down at all — ruling 4 already says nothing
+    // after it is fatal. Nor could the failure be reported: this writer *is* the
+    // error channel, and `logging.install` has not run yet, so `std.log` resolves
+    // to the same stderr a diagnostic about it would have to travel down.
+    //
+    // It is not lost from the process either. A failed drain consumes nothing,
+    // so whatever the buffer held it still holds — that half is observed, in
+    // "a broken error writer does not stop a first start that succeeded" below,
+    // which reads the retained warning back out of the same writer.
+    //
+    // What happens to those bytes afterwards is derived, not watched, and is
+    // labelled so deliberately. `Io.Writer.defaultFlush` drains while `end != 0`
+    // and `run`'s exit flush maps a failure to exit 1, so a stderr still broken
+    // at shutdown should carry the condition out in the exit code, and one that
+    // recovered should deliver the line late. No test drives `run` that far.
+    // Two limits come with the derivation: an empty buffer flushes clean and
+    // reports nothing at all, and a failure that recovers ends at exit 0 with a
+    // line the operator reads days after the start it describes.
+    diags.writeAll(r.err) catch {};
+    r.err.flush() catch {};
+    return result;
 }
 
 fn serve(r: cli.Runner, args: cli.RunArgs) !u8 {
@@ -141,18 +193,7 @@ fn serve(r: cli.Runner, args: cli.RunArgs) !u8 {
     defer config_db.close();
     _ = try migrations.migrate(&config_db);
 
-    {
-        // First run only: the file seeds an empty database and is ignored
-        // forever after. Its diagnostics are the operator's one chance to see
-        // why a config file was rejected, so they are printed like `check`
-        // prints them.
-        var diags: validate.Diagnostics = .init(gpa);
-        defer diags.deinit();
-        _ = bootstrap.bootstrap(io, gpa, &config_db, std.Io.Dir.cwd(), paths.config, &diags) catch |err| {
-            diags.writeAll(r.err) catch {};
-            return err;
-        };
-    }
+    _ = try seedFromFile(r, &config_db, std.Io.Dir.cwd(), paths.config);
 
     // Every string in `cfg` points into this arena, and the pool's endpoints,
     // the handler's records and the monitor's paths all keep such strings. It
@@ -717,7 +758,7 @@ const Upstreams = struct {
         http: *std.http.Client,
         bundle: *Certificate.Bundle,
         bundle_lock: *std.Io.RwLock,
-    ) (Allocator.Error || ConfigError)!Upstreams {
+    ) (Allocator.Error || error{NoUsableUpstreams})!Upstreams {
         var enabled: usize = 0;
         for (servers) |server| {
             if (server.enabled) enabled += 1;
@@ -748,7 +789,10 @@ const Upstreams = struct {
             if (!server.enabled) continue;
 
             const endpoint = transport.Endpoint.parse(server.url) catch {
-                log.warn("upstream '{s}' is not an https:// or tls:// endpoint; skipped", .{server.url});
+                log.warn(
+                    "upstream {f} is not an https:// or tls:// endpoint; skipped",
+                    .{safe_url.redactQuoted(server.url)},
+                );
                 continue;
             };
 
@@ -762,7 +806,10 @@ const Upstreams = struct {
                         self.doh_buf[base..][0..doh_request_buf_len],
                         self.doh_buf[base + doh_request_buf_len ..][0..doh_transfer_buf_len],
                     ) catch {
-                        log.warn("upstream '{s}' is not a usable DoH url; skipped", .{server.url});
+                        log.warn(
+                            "upstream {f} is not a usable DoH url; skipped",
+                            .{safe_url.redactQuoted(server.url)},
+                        );
                         continue;
                     };
                     doh_count += 1;
@@ -888,6 +935,280 @@ test "parseBind refuses a bind address of the wrong family" {
     try std.testing.expect(std.mem.containsAtLeast(u8, printed, 1, "is not an IPv4 address"));
 }
 
+test "run maps a rejected configuration to exit 2 and everything else to exit 1" {
+    try std.testing.expectEqual(cli.exit_check, failureExitCode(error.MissingDefaultGroup));
+    try std.testing.expectEqual(cli.exit_check, failureExitCode(error.ParseZon));
+    try std.testing.expectEqual(cli.exit_check, failureExitCode(error.NoUsableUpstreams));
+    try std.testing.expectEqual(cli.exit_check, failureExitCode(error.BadCertificate));
+    try std.testing.expectEqual(cli.exit_runtime, failureExitCode(error.AccessDenied));
+    try std.testing.expectEqual(cli.exit_runtime, failureExitCode(error.OutOfMemory));
+}
+
+test "run, check and import agree on a seed file with no default group" {
+    const config_import = @import("config/import.zig");
+
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    // The file from discrepancy D1: parseable, one upstream, no group named
+    // 'default'.
+    const source: [:0]const u8 =
+        \\.{
+        \\    .groups = .{ .{ .name = "kids" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\}
+    ;
+
+    // `run`: `serve` seeds through `bootstrap`, which is a wrapper over this
+    // exact call, so this is the error `run` classifies.
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    defer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+
+    var diags: validate.Diagnostics = .init(gpa);
+    defer diags.deinit();
+    try std.testing.expectError(
+        error.MissingDefaultGroup,
+        config_import.importSource(io, gpa, &database, source, .{}, &diags),
+    );
+    try std.testing.expectEqual(cli.exit_check, failureExitCode(error.MissingDefaultGroup));
+
+    // `import`: `cli.failureExitCode` reaches exit 2 by either route — the
+    // recorded failures, or the classification `run` just used.
+    try std.testing.expect(diags.failureCount() != 0);
+    try std.testing.expect(faults.isConfigFault(error.MissingDefaultGroup));
+
+    // `check`: the same file, through the code `nxdns check` runs.
+    var arena_state: std.heap.ArenaAllocator = .init(gpa);
+    defer arena_state.deinit();
+    const cfg = try std.zon.parse.fromSliceAlloc(model.Config, arena_state.allocator(), source, null, .{});
+
+    var out_buf: [2048]u8 = undefined;
+    var err_buf: [256]u8 = undefined;
+    var out: Writer = .fixed(&out_buf);
+    var err_writer: Writer = .fixed(&err_buf);
+    const r: cli.Runner = .{ .io = io, .gpa = gpa, .out = &out, .err = &err_writer };
+
+    try std.testing.expectEqual(cli.exit_check, try cli.checkConfig(r, cfg, false));
+}
+
+test "a configuration whose blocklist source is in no group imports and checks clean" {
+    const config_import = @import("config/import.zig");
+
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    // A source is created before it is attached — `docs/tutorial/first-run.md`
+    // POSTs the blocklist and then PUTs the group's sources — so an unattached
+    // source is a legal intermediate state on every write path. It warns; it
+    // never fails.
+    const source: [:0]const u8 =
+        \\.{
+        \\    .groups = .{ .{ .name = "default" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\    .blocklist_sources = .{ .{ .url = "https://lists.example/hosts.txt", .name = "ads" } },
+        \\}
+    ;
+
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    defer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+
+    var diags: validate.Diagnostics = .init(gpa);
+    defer diags.deinit();
+    try config_import.importSource(io, gpa, &database, source, .{}, &diags);
+    try std.testing.expectEqual(@as(usize, 0), diags.failureCount());
+    try std.testing.expectEqual(@as(usize, 1), diags.warningCount());
+
+    var arena_state: std.heap.ArenaAllocator = .init(gpa);
+    defer arena_state.deinit();
+    const cfg = try std.zon.parse.fromSliceAlloc(model.Config, arena_state.allocator(), source, null, .{});
+
+    var check_diags: validate.Diagnostics = .init(gpa);
+    defer check_diags.deinit();
+    // No error means no subcommand may reject it; the warning is report-only.
+    try validate.validate(cfg, &check_diags);
+    try std.testing.expectEqual(@as(usize, 0), check_diags.failureCount());
+    try std.testing.expectEqual(@as(usize, 1), check_diags.warningCount());
+}
+
+test "a first start that seeds from a file prints the warnings the file earned" {
+    // D5, second half. The seed file is read once in the life of a database, so
+    // a warning it earns is printed on that start or never. `run` printed
+    // diagnostics only when the file was rejected, which made a successful first
+    // start the one place the finding could not surface.
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    var tmp = std.testing.tmpDir(.{});
+    defer tmp.cleanup();
+    try tmp.dir.writeFile(io, .{ .sub_path = "config.zon", .data =
+        \\.{
+        \\    .groups = .{ .{ .name = "default" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\    .blocklist_sources = .{ .{ .url = "https://lists.example/hosts.txt", .name = "ads" } },
+        \\}
+    });
+
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    defer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+
+    // A buffered file writer, the shape `main` builds over stderr, and not
+    // `Writer.fixed`: a fixed writer's flush is a no-op, so it counts a line
+    // still sitting in the buffer as delivered. Reading the file back is the
+    // only way to ask what the operator can actually see.
+    var out_buf: [64]u8 = undefined;
+    var out: Writer = .fixed(&out_buf);
+    var err_file = try tmp.dir.createFile(io, "stderr.txt", .{});
+    defer err_file.close(io);
+    var err_buf: [4096]u8 = undefined;
+    var err_writer = err_file.writer(io, &err_buf);
+    const r: cli.Runner = .{ .io = io, .gpa = gpa, .out = &out, .err = &err_writer.interface };
+
+    // The file is valid, so the start succeeds and the database is seeded.
+    try std.testing.expectEqual(
+        bootstrap.Outcome.seeded,
+        try seedFromFile(r, &database, tmp.dir, "config.zon"),
+    );
+    try std.testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM upstreams"));
+
+    // Nothing flushes here on purpose. In production `serve` runs from this
+    // point until the service stops, so a line that has not reached the file by
+    // now is a line the operator does not get for days.
+    const printed = try tmp.dir.readFileAlloc(io, "stderr.txt", gpa, .limited(8192));
+    defer gpa.free(printed);
+
+    try std.testing.expect(std.mem.containsAtLeast(u8, printed, 1, "WARN blocklist_sources[0]: "));
+    try std.testing.expect(std.mem.containsAtLeast(u8, printed, 1, "belongs to no group"));
+    // A warning is not a rejection: nothing here claims the start failed.
+    try std.testing.expectEqual(@as(usize, 0), std.mem.count(u8, printed, "FAIL"));
+}
+
+/// A broken stderr, in the shape `main` builds: a buffered `File.Writer`, with
+/// its drain switched to the mode that fails. `Writer.fixed` cannot stand in —
+/// its flush is `noopFlush`, so it has no failure to report and its `written()`
+/// counts a line still sitting in the buffer as delivered.
+///
+/// The file stays empty for as long as the mode is `.failure`, which is what
+/// lets a test tell "the writer really failed" from "the writer worked".
+fn brokenErrWriter(io: std.Io, file: std.Io.File, buffer: []u8) std.Io.File.Writer {
+    var w = file.writer(io, buffer);
+    w.mode = .failure;
+    return w;
+}
+
+test "a broken error writer does not replace the reason a seed file was rejected" {
+    // The operator has to see why seeding failed, and a broken stderr is not
+    // that reason.
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    var tmp = std.testing.tmpDir(.{});
+    defer tmp.cleanup();
+    // No group named 'default': rejected, and it records a FAIL line on the way.
+    try tmp.dir.writeFile(io, .{ .sub_path = "config.zon", .data =
+        \\.{
+        \\    .groups = .{ .{ .name = "kids" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\}
+    });
+
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    defer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+
+    var out_buf: [64]u8 = undefined;
+    var out: Writer = .fixed(&out_buf);
+    var err_file = try tmp.dir.createFile(io, "stderr.txt", .{});
+    defer err_file.close(io);
+    // Eight bytes: no FAIL line fits, so `writeAll` must drain mid-line and is
+    // itself the call that fails. The test below covers the other discard, where
+    // the line fits and only the flush fails.
+    var err_buf: [8]u8 = undefined;
+    var err_writer = brokenErrWriter(io, err_file, &err_buf);
+    const r: cli.Runner = .{ .io = io, .gpa = gpa, .out = &out, .err = &err_writer.interface };
+
+    try std.testing.expectError(
+        error.MissingDefaultGroup,
+        seedFromFile(r, &database, tmp.dir, "config.zon"),
+    );
+
+    // Empty, so the writer did fail — without this the assertion above would
+    // hold just as well against a writer that worked.
+    const printed = try tmp.dir.readFileAlloc(io, "stderr.txt", gpa, .limited(8192));
+    defer gpa.free(printed);
+    try std.testing.expectEqual(@as(usize, 0), printed.len);
+}
+
+test "a broken error writer does not stop a first start that succeeded" {
+    // The call this file makes: a DNS server for a household does not refuse to
+    // resolve because stderr is broken. What it must not do is drop the warning
+    // on the floor, so the second half checks the buffer still holds it.
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    var tmp = std.testing.tmpDir(.{});
+    defer tmp.cleanup();
+    // Valid, and it earns one warning: the source belongs to no group.
+    try tmp.dir.writeFile(io, .{ .sub_path = "config.zon", .data =
+        \\.{
+        \\    .groups = .{ .{ .name = "default" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\    .blocklist_sources = .{ .{ .url = "https://lists.example/hosts.txt", .name = "ads" } },
+        \\}
+    });
+
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    defer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+
+    var out_buf: [64]u8 = undefined;
+    var out: Writer = .fixed(&out_buf);
+    var err_file = try tmp.dir.createFile(io, "stderr.txt", .{});
+    defer err_file.close(io);
+    // 4 KiB, the buffer `main` gives the real stderr writer: the warning fits,
+    // so `writeAll` succeeds into the buffer and the flush is what fails. That
+    // is the shape production hits.
+    var err_buf: [4096]u8 = undefined;
+    var err_writer = brokenErrWriter(io, err_file, &err_buf);
+    const r: cli.Runner = .{ .io = io, .gpa = gpa, .out = &out, .err = &err_writer.interface };
+
+    try std.testing.expectEqual(
+        bootstrap.Outcome.seeded,
+        try seedFromFile(r, &database, tmp.dir, "config.zon"),
+    );
+    try std.testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM upstreams"));
+
+    // Nothing reached the file, so the flush really did fail.
+    const undelivered = try tmp.dir.readFileAlloc(io, "stderr.txt", gpa, .limited(8192));
+    defer gpa.free(undelivered);
+    try std.testing.expectEqual(@as(usize, 0), undelivered.len);
+
+    // And the warning is still buffered rather than dropped: this is the same
+    // writer, and these are the bytes `run`'s exit flush meets on the way out.
+    err_writer.mode = .positional;
+    try err_writer.interface.flush();
+    const printed = try tmp.dir.readFileAlloc(io, "stderr.txt", gpa, .limited(8192));
+    defer gpa.free(printed);
+    try std.testing.expect(std.mem.containsAtLeast(u8, printed, 1, "WARN blocklist_sources[0]: "));
+}
+
 fn reportBind(r: cli.Runner, which: []const u8, addr: net.IpAddress, err: anyerror) anyerror {
     r.err.print("cannot bind {s} {f}: {s}\n", .{ which, addr, @errorName(err) }) catch {};
     return err;

src/cli.zig

@@ -20,9 +20,11 @@ const tls = std.crypto.tls;
 
 const app = @import("app.zig");
 const config_export = @import("config/export.zig");
+const faults = @import("config/faults.zig");
 const import = @import("config/import.zig");
 const model = @import("config/model.zig");
 const validate = @import("config/validate.zig");
+const cert_store = @import("server/cert_store.zig");
 const db = @import("storage/db.zig");
 const migrations = @import("storage/migrations.zig");
 const querylog_schema = @import("storage/querylog_schema.zig");
@@ -30,6 +32,7 @@ const doh_client = @import("upstream/doh_client.zig");
 const dot_client = @import("upstream/dot_client.zig");
 const pool = @import("upstream/pool.zig");
 const transport = @import("upstream/transport.zig");
+const safe_url = @import("safe_url.zig");
 const version = @import("version.zig");
 
 pub const exit_ok: u8 = 0;
@@ -436,7 +439,7 @@ pub fn runImport(r: Runner, args: ImportArgs) u8 {
         // should need one run to see the whole list.
         diags.writeAll(r.err) catch {};
         r.err.print("import failed: {s}\n", .{@errorName(e)}) catch {};
-        return finish(r, failureExitCode(e, diags.problems.items.len));
+        return finish(r, failureExitCode(e, diags.failureCount()));
     };
     return finish(r, exit_ok);
 }
@@ -458,32 +461,42 @@ fn importImpl(r: Runner, args: ImportArgs, diags: *validate.Diagnostics) !void {
         .{ .force = args.force },
         diags,
     );
+
+    // A configuration that imports cleanly can still have recorded warnings — a
+    // blocklist source in no group is the one that found this. `validate`
+    // returns nothing for a warning, so printing diagnostics on the failure path
+    // alone made `import` the command that read the finding and threw it away,
+    // while `check` printed it from the same file. Only warnings can be here:
+    // any recorded failure returned above.
+    try diags.writeAll(r.out);
     try r.out.print("imported {s}\n", .{args.file});
 }
 
 /// A configuration the operator can fix exits 2; everything else is a runtime
 /// failure. `validate` records a diagnostic for every error it returns and then
-/// returns the first one, so a non-empty diagnostics list is the reliable
-/// discriminator; the named errors below are the config faults that never reach
-/// the validator.
+/// returns the first one, so a recorded failure is the reliable discriminator;
+/// `config/faults.zig` classifies the errors that never reach the validator.
+/// That file is the only list — this function keeps none of its own, which is
+/// what stops `run`, `check` and `import` drifting apart again (D1).
 ///
-/// `error.OutOfMemory` is matched first, before the list is consulted. Both
-/// recording paths — `validate` and import's per-line rendering of a ZON syntax
-/// error — add one problem at a time and can run out of memory partway, which
-/// leaves problems recorded for a run whose real outcome is a resource failure.
-/// A partial report is not a verdict on the configuration, so the runtime exit
-/// code wins.
-fn failureExitCode(e: anyerror, problems: usize) u8 {
+/// `error.DatabaseNotEmpty` is the one exception, and it is deliberate: it
+/// reports the state of the database rather than the content of a file, so it
+/// is not a configuration fault, and `import` alone decides it is exit 2.
+///
+/// `error.OutOfMemory` is matched first, before anything else is consulted.
+/// Both recording paths — `validate` and import's per-line rendering of a ZON
+/// syntax error — add one problem at a time and can run out of memory partway,
+/// which leaves problems recorded for a run whose real outcome is a resource
+/// failure. A partial report is not a verdict on the configuration, so the
+/// runtime exit code wins.
+///
+/// `failures` counts recorded failures, never warnings: a warning never changes
+/// an exit code (F-b).
+fn failureExitCode(e: anyerror, failures: usize) u8 {
     if (e == error.OutOfMemory) return exit_runtime;
-    if (problems != 0) return exit_check;
-    return switch (e) {
-        error.DatabaseNotEmpty,
-        error.ConfigTooLarge,
-        error.ParseZon,
-        error.PasswordAndHashBothSet,
-        => exit_check,
-        else => exit_runtime,
-    };
+    if (failures != 0) return exit_check;
+    if (e == error.DatabaseNotEmpty) return exit_check;
+    return if (faults.isConfigFault(e)) exit_check else exit_runtime;
 }
 
 // ---------------------------------------------------------------------------
@@ -520,21 +533,10 @@ fn checkImpl(r: Runner, args: CheckArgs, probe: bool) !u8 {
         return checkFile(r, arena, args.paths.config, probe);
     }
 
-    const config_db_path = try std.fs.path.join(arena, &.{ args.paths.data_dir, config_db_name });
+    const config_db_path = try std.fs.path.joinZ(arena, &.{ args.paths.data_dir, config_db_name });
     if (try pathExists(r.io, config_db_path)) {
         try r.out.print("checking database {s}\n", .{config_db_path});
-
-        var data = try DataDir.open(r.io, r.gpa, args.paths.data_dir, false);
-        defer data.close(r.io, r.gpa);
-
-        // A `check` on a database one schema version behind must still work,
-        // which is what an operator runs right after an upgrade.
-        var database = try data.openConfigDb(r.io);
-        defer database.close();
-        _ = try migrations.migrate(&database);
-
-        const cfg = try config_export.readConfig(&database, arena);
-        return checkConfig(r, cfg, probe);
+        return checkDatabase(r, arena, config_db_path, probe);
     }
 
     if (try pathExists(r.io, args.paths.config)) {
@@ -550,6 +552,121 @@ fn checkImpl(r: Runner, args: CheckArgs, probe: bool) !u8 {
     return exit_check;
 }
 
+/// `check` reads `config.db` and writes nothing to it (F-c): no create, no
+/// chmod, no `applyPragmas` — which is what would turn WAL on and leave
+/// `config.db-wal` and `config.db-shm` behind — and above all no `migrate`. A
+/// database behind this binary's schema is reported here; upgrading it is
+/// `nxdns run`'s job, and a command that claims to validate without writing may
+/// not commit a schema step.
+///
+/// `.immutable` is the enforcement, not a convention: SQLite refuses every
+/// statement that would write, and the mode's `immutable=1` also stops the
+/// pager building a wal-index, so no `config.db-wal` and no `config.db-shm`
+/// appear beside the file. `.read_only` alone creates both and cannot delete
+/// them on close, which is how a "validate without writing" command left two
+/// files behind.
+fn checkDatabase(r: Runner, arena: Allocator, path: [:0]const u8, probe: bool) !u8 {
+    var database = db.Db.open(path, .{ .mode = .{ .immutable = r.io } }) catch |e| switch (e) {
+        error.OutOfMemory => return error.OutOfMemory,
+        error.WalPending => return walPendingFailure(r, path),
+        else => return unreadable(r, path, e),
+    };
+    defer database.close();
+
+    const reading = try readDatabase(&database, arena);
+
+    // `immutable=1` takes no lock — that is what keeps it from building a
+    // wal-index and leaving two sidecars behind — so a writer was free to append
+    // to the log or checkpoint into the main file for the whole of the read
+    // above. Proved before one word of it is reported: a stale answer and a torn
+    // read both look exactly like an ordinary finding, which is how this failure
+    // stays invisible.
+    database.verifyImmutable() catch |e| switch (e) {
+        error.OutOfMemory => return error.OutOfMemory,
+        error.WalPending => return walPendingFailure(r, path),
+        else => return unreadable(r, path, e),
+    };
+
+    switch (reading) {
+        .version_unreadable => |e| {
+            try r.out.print("FAIL {s}: the schema version cannot be read ({s})\n", .{ path, @errorName(e) });
+            return exit_check;
+        },
+        // Naming both numbers is the point: "at 1, expects 2" tells an operator
+        // to run `nxdns run`, where a bare SQLite complaint about a missing
+        // column tells them nothing.
+        .version_mismatch => |stamped| {
+            const fix = if (stamped < migrations.target_version)
+                "`nxdns run` migrates it, `check` will not"
+            else
+                "it was written by a newer nxdns";
+            try r.out.print(
+                "FAIL {s}: schema version {d}, this nxdns expects {d}; {s}\n",
+                .{ path, stamped, migrations.target_version, fix },
+            );
+            return exit_check;
+        },
+        .config_unreadable => {
+            // The schema version is already known good, so whatever this is,
+            // the SQLite message is the only thing that narrows it down.
+            // `verifyImmutable` makes no SQLite call, so this is still the
+            // message from the read.
+            var buf: [256]u8 = undefined;
+            try r.out.print("FAIL {s}: cannot be read ({s})\n", .{ path, database.lastError(&buf) });
+            return exit_check;
+        },
+        .config => |cfg| return checkConfig(r, cfg, probe),
+    }
+}
+
+/// Everything read out of `config.db`, held rather than reported, because a
+/// value from an unlocked read is only worth reporting once `verifyImmutable`
+/// has said the files it came from stood still. A wrong schema-version line is
+/// as misleading as a wrong setting.
+const DbReading = union(enum) {
+    config: model.Config,
+    version_unreadable: anyerror,
+    version_mismatch: u32,
+    config_unreadable,
+};
+
+/// Reads only, so an immutable handle serves it.
+fn readDatabase(database: *db.Db, arena: Allocator) error{OutOfMemory}!DbReading {
+    const stamped = migrations.readVersion(database) catch |e| switch (e) {
+        error.OutOfMemory => return error.OutOfMemory,
+        else => return .{ .version_unreadable = e },
+    };
+    if (stamped != migrations.target_version) return .{ .version_mismatch = stamped };
+
+    const cfg = config_export.readConfig(database, arena) catch |e| switch (e) {
+        error.OutOfMemory => return error.OutOfMemory,
+        else => return .config_unreadable,
+    };
+    return .{ .config = cfg };
+}
+
+/// One wording for `error.WalPending`, whether the log was already there when
+/// the read opened or arrived while it ran: from the operator's side those are
+/// the same situation, a writer holding changes this read cannot see.
+///
+/// `immutable=1` ignores the write-ahead log, so the newest committed settings
+/// would be invisible and `check` would quietly grade the older ones in the main
+/// file. The guard is deliberately conservative — a live writer, an interrupted
+/// process and a checkpointed log that was simply kept all look the same from
+/// outside — so the line says what to do and does not claim anything is damaged.
+fn walPendingFailure(r: Runner, path: []const u8) !u8 {
+    try r.out.print(
+        "FAIL {s}: uncheckpointed changes are waiting in {s}{s}, and reading without writing would answer from the older settings in the main file; `nxdns run` applies them. A running nxdns normally holds this log, which is the usual reason to see this line.\n",
+        .{ path, path, db.wal_suffix },
+    );
+    return exit_check;
+}
+
+fn unreadable(r: Runner, path: []const u8, e: anyerror) !u8 {
+    try r.out.print("FAIL {s}: cannot be opened for reading ({s})\n", .{ path, @errorName(e) });
+    return exit_check;
+}
+
 fn pathExists(io: std.Io, path: []const u8) std.Io.Dir.AccessError!bool {
     std.Io.Dir.cwd().access(io, path, .{}) catch |e| switch (e) {
         error.FileNotFound => return false,
@@ -583,6 +700,19 @@ fn checkFile(r: Runner, arena: Allocator, path: []const u8, probe: bool) !u8 {
             try r.out.print("FAIL {s}: larger than {d} bytes\n", .{ path, import.max_config_bytes });
             return exit_check;
         },
+        // D4: a named file that is missing or unreadable is the same
+        // operator-fixable condition as one that fails to parse, so it is
+        // reported as a finding rather than escaping as a runtime failure. The
+        // implicit path already exits 2 when it finds nothing to check; naming
+        // the file must not change the code.
+        error.FileNotFound => {
+            try r.out.print("FAIL {s}: no such file\n", .{path});
+            return exit_check;
+        },
+        error.AccessDenied, error.PermissionDenied => {
+            try r.out.print("FAIL {s}: not readable\n", .{path});
+            return exit_check;
+        },
         else => |other| return other,
     };
 
@@ -602,48 +732,83 @@ fn checkFile(r: Runner, arena: Allocator, path: []const u8, probe: bool) !u8 {
     return checkConfig(r, cfg, probe);
 }
 
+/// What one part of `check` found. Failures set exit 2; warnings are printed,
+/// counted for the summary, and never change an exit code (F-b) — the service
+/// starts either way.
+const Tally = struct {
+    failures: usize = 0,
+    warnings: usize = 0,
+
+    fn plus(self: Tally, other: Tally) Tally {
+        return .{
+            .failures = self.failures + other.failures,
+            .warnings = self.warnings + other.warnings,
+        };
+    }
+};
+
 /// The half of `check` that has a `Config` already: validate, report every
 /// diagnostic, then the certificate and upstream checks. With TLS disabled and
 /// `probe` false it touches neither the filesystem nor the network, which is
 /// what makes it unit-testable.
+///
+/// The summary never contradicts the lines above it (D2): a run that printed
+/// WARN lines says so instead of claiming no problems were found, and still
+/// exits 0.
 pub fn checkConfig(r: Runner, cfg: model.Config, probe: bool) !u8 {
     var diags: validate.Diagnostics = .init(r.gpa);
     defer diags.deinit();
 
-    // The returned error is `problems[0].err` — one of the lines about to be
-    // printed — so it carries nothing the report does not. Only an allocation
-    // failure means the report itself is incomplete.
+    // The returned error is the first recorded failure — one of the lines about
+    // to be printed — so it carries nothing the report does not. Only an
+    // allocation failure means the report itself is incomplete.
     validate.validate(cfg, &diags) catch |e| switch (e) {
         error.OutOfMemory => return error.OutOfMemory,
         else => {},
     };
+    // `writeAll` prints the "FAIL "/"WARN " prefix itself; the lines below add
+    // their own because they are not diagnostics.
     try diags.writeAll(r.out);
 
-    var failures = diags.problems.items.len;
-    failures += try checkCertificates(r, cfg);
-    if (probe) failures += try probeUpstreams(r, cfg);
+    var tally: Tally = .{ .failures = diags.failureCount(), .warnings = diags.warningCount() };
+    tally = tally.plus(try checkCertificates(r, cfg));
+    if (probe) tally.failures += try probeUpstreams(r, cfg);
 
-    if (failures != 0) return exit_check;
+    if (tally.failures != 0) return exit_check;
+    if (tally.warnings != 0) {
+        try r.out.print("OK: no failures found, {d} warning{s}\n", .{
+            tally.warnings,
+            if (tally.warnings == 1) "" else "s",
+        });
+        return exit_ok;
+    }
     try r.out.writeAll("OK: no problems found\n");
     return exit_ok;
 }
 
-fn checkCertificates(r: Runner, cfg: model.Config) !usize {
-    return try checkTlsFiles(r, cfg.doh_server, "doh_server") +
-        try checkTlsFiles(r, cfg.dot_server, "dot_server");
+fn checkCertificates(r: Runner, cfg: model.Config) !Tally {
+    const doh = try checkTlsFiles(r, cfg.doh_server, "doh_server");
+    const dot = try checkTlsFiles(r, cfg.dot_server, "dot_server");
+    return doh.plus(dot);
 }
 
-/// An unreadable certificate or key fails the run; a key readable by anyone
-/// beyond its owner is a warning (PLAN §19) and leaves the exit code alone,
-/// because the service still starts.
-fn checkTlsFiles(r: Runner, endpoint: model.TlsEndpoint, comptime section: []const u8) !usize {
-    if (!endpoint.enabled) return 0;
-    var failures: usize = 0;
-
-    if (!try pathReadable(r.io, endpoint.cert_path)) {
-        try r.out.print("FAIL " ++ section ++ ".cert_path: '{s}' is not readable\n", .{endpoint.cert_path});
-        failures += 1;
-    }
+/// Proves the pair rather than the paths (D3). `CertStore.init` is the load the
+/// listeners boot with: it reads both PEM files and builds a
+/// `tls_server.ServerContext`, which is where Mbed TLS parses the chain, parses
+/// the key and checks that the key belongs to the leaf. Testing readability
+/// alone let `check` exit 0 on a certificate and key that do not pair, seconds
+/// before `run` exited 2 on `BadCertificate`.
+///
+/// No listener is bound and nothing is published: the context is built and
+/// freed. `alpn` is null because ALPN is negotiated per connection and plays no
+/// part in loading a pair; every other input is the server's.
+///
+/// A key readable by anyone beyond its owner stays a warning (PLAN §19) and is
+/// reported even when the pair itself fails, because it is a separate finding
+/// about a file that exists.
+fn checkTlsFiles(r: Runner, endpoint: model.TlsEndpoint, comptime section: []const u8) !Tally {
+    if (!endpoint.enabled) return .{};
+    var tally: Tally = .{};
 
     if (try pathReadable(r.io, endpoint.key_path)) {
         const stat = try std.Io.Dir.cwd().statFile(r.io, endpoint.key_path, .{});
@@ -653,13 +818,46 @@ fn checkTlsFiles(r: Runner, endpoint: model.TlsEndpoint, comptime section: []con
                 "WARN " ++ section ++ ".key_path: '{s}' is mode {o}; a TLS key must be readable by its owner only\n",
                 .{ endpoint.key_path, mode },
             );
+            tally.warnings += 1;
         }
-    } else {
-        try r.out.print("FAIL " ++ section ++ ".key_path: '{s}' is not readable\n", .{endpoint.key_path});
-        failures += 1;
     }
 
-    return failures;
+    var store = cert_store.CertStore.init(
+        r.gpa,
+        r.io,
+        endpoint.cert_path,
+        endpoint.key_path,
+        null,
+    ) catch |e| {
+        const at: struct { field: []const u8, path: []const u8 } = switch (e) {
+            error.OutOfMemory => return error.OutOfMemory,
+            // Not a verdict on the configuration: the platform's entropy source
+            // failed, which is the same failure `run` would hit.
+            error.EntropyFailed => return error.EntropyFailed,
+            error.CertUnreadable,
+            error.CertTooLarge,
+            error.CertParse,
+            // Mbed TLS rejected the server configuration built from this pair,
+            // so the certificate is what the operator has to look at.
+            error.ConfigFailed,
+            => .{ .field = "cert_path", .path = endpoint.cert_path },
+            error.KeyUnreadable,
+            error.KeyTooLarge,
+            error.KeyParse,
+            error.KeyMismatch,
+            => .{ .field = "key_path", .path = endpoint.key_path },
+        };
+        try r.out.print("FAIL " ++ section ++ ".{s}: '{s}': {s}\n", .{
+            at.field,
+            at.path,
+            cert_store.humanMessage(e),
+        });
+        tally.failures += 1;
+        return tally;
+    };
+    store.deinit(r.io);
+
+    return tally;
 }
 
 /// One `Pool` per upstream, never one pool over all of them. The pool's job is
@@ -697,11 +895,19 @@ fn probeUpstreams(r: Runner, cfg: model.Config) !usize {
     // backoff, so the value only has to be a value.
     const seed: u64 = @truncate(@as(u96, @bitCast(std.Io.Clock.real.now(r.io).nanoseconds)));
 
-    for (cfg.upstreams) |server| {
+    // Every line below names the upstream by its index into the configuration
+    // as well as by its redacted url, and the index is the config index rather
+    // than a count of the upstreams probed — a disabled entry still occupies
+    // one. The url alone no longer identifies an entry: `safe_url.redact` drops
+    // the path, and two upstreams on one host commonly differ only there (a
+    // NextDNS profile is `https://dns.nextdns.io/<profile>`). `upstreams[N]` is
+    // the same name `config/validate.zig` gives the entry, so a FAIL line here
+    // and a FAIL line from the validator point at the same place.
+    for (cfg.upstreams, 0..) |server, i| {
         if (!server.enabled) continue;
 
         const endpoint = transport.Endpoint.parse(server.url) catch {
-            try r.out.print("FAIL {s}: not an https:// or tls:// endpoint\n", .{server.url});
+            try r.out.print("FAIL upstreams[{d}] {f}: not an https:// or tls:// endpoint\n", .{ i, safe_url.redactQuoted(server.url) });
             failures += 1;
             continue;
         };
@@ -713,7 +919,7 @@ fn probeUpstreams(r: Runner, cfg: model.Config) !usize {
         const client: transport.Client = switch (endpoint.scheme) {
             .doh => doh: {
                 doh = doh_client.DohClient.init(&http, endpoint, &request_buf, &transfer_buf) catch {
-                    try r.out.print("FAIL {s}: not a usable DoH url\n", .{server.url});
+                    try r.out.print("FAIL upstreams[{d}] {f}: not a usable DoH url\n", .{ i, safe_url.redactQuoted(server.url) });
                     failures += 1;
                     continue;
                 };
@@ -740,7 +946,7 @@ fn probeUpstreams(r: Runner, cfg: model.Config) !usize {
         var single: pool.Pool = .init(&entries, .{}, attempt_timeout, seed);
 
         if (single.exchange(r.io, probe_query, response_buf)) |_| {
-            try r.out.print("OK {s}\n", .{server.url});
+            try r.out.print("OK upstreams[{d}] {f}\n", .{ i, safe_url.redact(server.url) });
         } else |_| {
             // The concrete cause lives in the entry's health, which is where the
             // pool put it; `@errorName` of the pool's return value would only
@@ -748,7 +954,7 @@ fn probeUpstreams(r: Runner, cfg: model.Config) !usize {
             var snapshots: [1]pool.Snapshot = undefined;
             const taken = try single.snapshot(r.io, &snapshots);
             const detail = if (taken == 1) snapshots[0].last_error else "no detail recorded";
-            try r.out.print("FAIL {s}: {s}\n", .{ server.url, detail });
+            try r.out.print("FAIL upstreams[{d}] {f}: {s}\n", .{ i, safe_url.redactQuoted(server.url), detail });
             failures += 1;
         }
     }
@@ -1019,6 +1225,391 @@ test "an allocation failure while recording diagnostics exits 1, not 2" {
     try testing.expect(saw_partial_report);
 }
 
-// `runCheck` against real paths, `runExport` and `runImport` all need a real
-// data directory, and the upstream probe leaves the machine. Those cases are
-// S7's: `src/storage/storage_integration_test.zig` cases 20-22.
+test "failureExitCode keeps no list of its own and classifies through config/faults.zig" {
+    // D1: every one of these reached `import` from a rejected seed file and was
+    // classified as a runtime failure by the private list this function used to
+    // carry, while `check` called the same file a configuration fault.
+    try testing.expectEqual(exit_check, failureExitCode(error.MissingDefaultGroup, 0));
+    try testing.expectEqual(exit_check, failureExitCode(error.NoUpstreams, 0));
+    try testing.expectEqual(exit_check, failureExitCode(error.BadUpstreamUrl, 0));
+    try testing.expectEqual(exit_check, failureExitCode(error.NoUsableUpstreams, 0));
+    try testing.expectEqual(exit_check, failureExitCode(error.BadCertificate, 0));
+
+    // Every member of the shared classification, so a variant added to
+    // `ValidateError` cannot exit 1 from `import` while exiting 2 from `run`.
+    inline for (@typeInfo(validate.ValidateError).error_set.?) |member| {
+        const err = @field(anyerror, member.name);
+        const expected: u8 = if (faults.isConfigFault(err)) exit_check else exit_runtime;
+        try testing.expectEqual(expected, failureExitCode(err, 0));
+    }
+
+    // The one config-shaped exit 2 `cli` still decides for itself: it reports
+    // the state of the database, not the content of a file.
+    try testing.expect(!faults.isConfigFault(error.DatabaseNotEmpty));
+    try testing.expectEqual(exit_check, failureExitCode(error.DatabaseNotEmpty, 0));
+}
+
+const fixtures = @import("test_fixtures");
+
+/// A tmp directory holding the fixture PEM pair and a data directory, addressed
+/// by cwd-relative paths the same way an operator's configuration names them.
+/// Must not move after `init`: the slices point into the buffers.
+const CheckEnv = struct {
+    tmp: testing.TmpDir,
+    cert_path_buf: [160]u8,
+    key_path_buf: [160]u8,
+    data_dir_buf: [160]u8,
+    missing_path_buf: [160]u8,
+    cert_path: []const u8,
+    key_path: []const u8,
+    data_dir: []const u8,
+    /// A path inside the tmp directory that is never created.
+    missing_path: []const u8,
+
+    fn init(env: *CheckEnv) !void {
+        env.tmp = testing.tmpDir(.{});
+        errdefer env.tmp.cleanup();
+        env.cert_path = try env.path(&env.cert_path_buf, "cert.pem");
+        env.key_path = try env.path(&env.key_path_buf, "key.pem");
+        env.data_dir = try env.path(&env.data_dir_buf, "data");
+        env.missing_path = try env.path(&env.missing_path_buf, "no-such-config.zon");
+    }
+
+    fn deinit(env: *CheckEnv) void {
+        env.tmp.cleanup();
+    }
+
+    fn path(env: *const CheckEnv, buf: []u8, name: []const u8) ![]const u8 {
+        return std.fmt.bufPrint(buf, ".zig-cache/tmp/{s}/{s}", .{ env.tmp.sub_path, name });
+    }
+
+    /// The key lands at 0600, so only a test that asks for the permission
+    /// warning gets one.
+    fn writePair(env: *CheckEnv, io: std.Io, key_pem: []const u8) !void {
+        try env.tmp.dir.writeFile(io, .{ .sub_path = "cert.pem", .data = fixtures.cert_pem });
+        try env.tmp.dir.writeFile(io, .{ .sub_path = "key.pem", .data = key_pem });
+        try env.tmp.dir.setFilePermissions(io, "key.pem", .fromMode(0o600), .{});
+    }
+
+    /// Valid but for whatever the test broke about the TLS pair.
+    fn tlsConfig(env: *const CheckEnv) model.Config {
+        return .{
+            .groups = &.{.{ .name = "default" }},
+            .upstreams = &.{.{ .url = "https://dns.example/dns-query" }},
+            .doh_server = .{
+                .enabled = true,
+                .cert_path = env.cert_path,
+                .key_path = env.key_path,
+            },
+        };
+    }
+
+    fn expectAbsent(env: *CheckEnv, io: std.Io, name: []const u8) !void {
+        env.tmp.dir.access(io, name, .{}) catch |e| switch (e) {
+            error.FileNotFound => return,
+            else => |other| return other,
+        };
+        std.debug.print("expected '{s}' not to exist\n", .{name});
+        return error.TestUnexpectedResult;
+    }
+};
+
+test "check fails a certificate and a key that do not pair" {
+    // D3: readability alone passed this configuration, and `run` then exited 2
+    // on `BadCertificate` seconds later.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+    try env.writePair(r.io, fixtures.mismatched_key_pem);
+
+    try testing.expectEqual(exit_check, try checkConfig(r, env.tlsConfig(), false));
+
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "FAIL doh_server.key_path"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "does not belong to the certificate"));
+    try testing.expectEqual(@as(usize, 0), std.mem.count(u8, text, "OK:"));
+}
+
+test "check fails a certificate file that does not parse" {
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+    try env.writePair(r.io, fixtures.key_pem);
+    try env.tmp.dir.writeFile(r.io, .{ .sub_path = "cert.pem", .data = "not a certificate\n" });
+
+    try testing.expectEqual(exit_check, try checkConfig(r, env.tlsConfig(), false));
+
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "FAIL doh_server.cert_path"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "could not be parsed"));
+}
+
+test "check passes a certificate and a key that pair" {
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+    try env.writePair(r.io, fixtures.key_pem);
+
+    try testing.expectEqual(exit_ok, try checkConfig(r, env.tlsConfig(), false));
+    try testing.expectEqualStrings("OK: no problems found\n", captured.out.written());
+}
+
+test "a check whose only findings are warnings exits 0 and says so" {
+    // D2: the summary used to read "OK: no problems found" directly under the
+    // WARN line it was contradicting.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+    try env.writePair(r.io, fixtures.key_pem);
+    try env.tmp.dir.setFilePermissions(r.io, "key.pem", .fromMode(0o644), .{});
+
+    // A warning never changes an exit code: the service still starts.
+    try testing.expectEqual(exit_ok, try checkConfig(r, env.tlsConfig(), false));
+
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "WARN doh_server.key_path"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "OK: no failures found, 1 warning\n"));
+    try testing.expectEqual(@as(usize, 0), std.mem.count(u8, text, "no problems found"));
+}
+
+test "check --config naming a missing file is a reported failure at exit 2" {
+    // D4: this escaped `checkImpl` as `check failed: FileNotFound` at exit 1,
+    // while the implicit path exits 2 for the same operator-fixable condition.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    const code = runCheck(r, .{
+        .paths = .{ .config = env.missing_path },
+        .config_explicit = true,
+    }, false);
+
+    try testing.expectEqual(exit_check, code);
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "FAIL"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "no such file"));
+    try testing.expectEqualStrings("", captured.err.written());
+}
+
+test "check reads config.db without writing to it" {
+    // D6: the database branch opened read/write, chmod'ed 0600, turned WAL on —
+    // which is what creates the two sidecars — and committed migration steps,
+    // from a command whose contract is that it validates without writing.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    {
+        var data = try DataDir.open(r.io, r.gpa, env.data_dir, true);
+        defer data.close(r.io, r.gpa);
+        var database = try data.openConfigDb(r.io);
+        defer database.close();
+        _ = try migrations.migrate(&database);
+    }
+
+    // A mode `openConfigDb` would overwrite, so its chmod cannot hide.
+    try env.tmp.dir.setFilePermissions(r.io, "data/config.db", .fromMode(0o644), .{});
+    const before = try env.tmp.dir.statFile(r.io, "data/config.db", .{});
+
+    _ = runCheck(r, .{ .paths = .{ .data_dir = env.data_dir } }, false);
+    try testing.expect(std.mem.containsAtLeast(u8, captured.out.written(), 1, "checking database"));
+
+    // Byte for byte the database the writer left, at the mode the writer left:
+    // no migration step committed, no chmod, no `PRAGMA journal_mode`.
+    const after = try env.tmp.dir.statFile(r.io, "data/config.db", .{});
+    try testing.expectEqual(before.size, after.size);
+    try testing.expectEqual(before.mtime.nanoseconds, after.mtime.nanoseconds);
+    try testing.expectEqual(
+        @as(@TypeOf(after.permissions.toMode()), 0o644),
+        after.permissions.toMode() & 0o777,
+    );
+
+    // Nothing beside it either. A `.read_only` open recreates the wal-index of
+    // a database whose header says WAL and cannot delete it on close, which
+    // left `config.db-wal` and `config.db-shm` behind; `.immutable` builds no
+    // wal-index at all.
+    try env.expectAbsent(r.io, "data/config.db-wal");
+    try env.expectAbsent(r.io, "data/config.db-shm");
+}
+
+test "check refuses a database whose write-ahead log still holds changes" {
+    // `immutable=1` ignores the log, so grading the main file would silently
+    // report settings the operator already replaced.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    {
+        var data = try DataDir.open(r.io, r.gpa, env.data_dir, true);
+        defer data.close(r.io, r.gpa);
+        var database = try data.openConfigDb(r.io);
+        defer database.close();
+        _ = try migrations.migrate(&database);
+    }
+    // Bytes are all the guard reads, and it never gets as far as opening this
+    // as a log: `db.Db.open` refuses on the size alone.
+    try env.tmp.dir.writeFile(r.io, .{
+        .sub_path = "data/config.db" ++ db.wal_suffix,
+        .data = "uncheckpointed frames",
+    });
+
+    const code = runCheck(r, .{ .paths = .{ .data_dir = env.data_dir } }, false);
+
+    try testing.expectEqual(exit_check, code);
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "FAIL"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "uncheckpointed changes"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "config.db" ++ db.wal_suffix));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "nxdns run"));
+    // Not a claim about damage: an operator reading this must not reach for a
+    // recovery tool.
+    try testing.expectEqual(@as(usize, 0), std.mem.count(u8, text, "corrupt"));
+}
+
+test "check reports a database behind the schema rather than migrating it" {
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    // Zero bytes is a valid, empty SQLite database: schema version 0, which is
+    // exactly what an upgrade leaves behind when a step has not run yet.
+    _ = try std.Io.Dir.cwd().createDirPathStatus(r.io, env.data_dir, .fromMode(0o700));
+    try env.tmp.dir.writeFile(r.io, .{ .sub_path = "data/config.db", .data = "" });
+
+    const code = runCheck(r, .{ .paths = .{ .data_dir = env.data_dir } }, false);
+
+    try testing.expectEqual(exit_check, code);
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "FAIL"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "nxdns run"));
+
+    // Both real numbers, not a generic SQLite complaint: an empty database is
+    // at version 0, and the expected one comes from the migration list rather
+    // than a literal, so adding a step cannot make this line lie.
+    var expected_buf: [96]u8 = undefined;
+    const expected = try std.fmt.bufPrint(
+        &expected_buf,
+        "schema version 0, this nxdns expects {d};",
+        .{migrations.target_version},
+    );
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, expected));
+
+    // Not upgraded: the schema `migrate` would have committed is still absent.
+    const after = try env.tmp.dir.statFile(r.io, "data/config.db", .{});
+    try testing.expectEqual(@as(u64, 0), after.size);
+    try env.expectAbsent(r.io, "data/config.db-wal");
+    try env.expectAbsent(r.io, "data/config.db-shm");
+}
+
+test "the upstream probe redacts a url it cannot parse, without leaving the machine" {
+    // No socket is opened on this branch: `Endpoint.parse` rejects the `@`
+    // before the loop builds a client, so the leak is reachable in a required
+    // test rather than only behind a live probe.
+    //
+    // It is also the only probe branch a credential can reach. `Endpoint.parse`
+    // refuses `@`, `?` and `#` in the authority and `?`/`#` in the path, so a
+    // url carrying userinfo or a query never gets as far as `DohClient.init`
+    // and its "not a usable DoH url" line. That line's redaction is defensive.
+    //
+    // The two upstreams are the NextDNS shape, where the profile id in the path
+    // is the account credential and is the only thing telling two entries
+    // apart. Both halves of the contract are asserted at once: neither the
+    // userinfo nor the profile id may reach stdout, and what is left has to
+    // still say which of the two entries the operator must go and fix. The
+    // disabled entry ahead of them is there because the index has to be the
+    // index into `upstreams`, not a count of the entries probed.
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    const cfg: model.Config = .{
+        .groups = &.{.{ .name = "default" }},
+        .upstreams = &.{
+            .{ .url = "https://dns.example/dns-query", .enabled = false },
+            .{ .url = "https://lists:hunter2@dns.nextdns.io/abcd12" },
+            .{ .url = "https://lists:hunter2@dns.nextdns.io/efgh34" },
+        },
+    };
+
+    try testing.expectEqual(@as(usize, 2), try probeUpstreams(r, cfg));
+
+    const text = captured.out.written();
+    try testing.expectEqualStrings(
+        "FAIL upstreams[1] 'https://dns.nextdns.io': not an https:// or tls:// endpoint\n" ++
+            "FAIL upstreams[2] 'https://dns.nextdns.io': not an https:// or tls:// endpoint\n",
+        text,
+    );
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "hunter2"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "abcd12"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "efgh34"));
+}
+
+test "a successful import prints the warnings the file earned" {
+    // D5, second half. `check` printed this WARN and `import` recorded it and
+    // threw it away, because diagnostics were written on the failure path only.
+    // The operator who imported the file learned nothing about the source they
+    // had just added, which blocks nothing until a group links it.
+    var env: CheckEnv = undefined;
+    try env.init();
+    defer env.deinit();
+
+    var captured: Captured = .init(testing.allocator);
+    defer captured.deinit();
+    const r = captured.runner();
+
+    try env.tmp.dir.writeFile(r.io, .{ .sub_path = "config.zon", .data =
+        \\.{
+        \\    .groups = .{ .{ .name = "default" } },
+        \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+        \\    .blocklist_sources = .{ .{ .url = "https://lists.example/hosts.txt", .name = "ads" } },
+        \\}
+    });
+
+    var file_buf: [160]u8 = undefined;
+    const file = try env.path(&file_buf, "config.zon");
+    const code = runImport(r, .{ .paths = .{ .data_dir = env.data_dir }, .file = file });
+
+    // A warning never changes an exit code, and the import really happened.
+    try testing.expectEqual(exit_ok, code);
+    const text = captured.out.written();
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "WARN blocklist_sources[0]: "));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "belongs to no group"));
+    try testing.expect(std.mem.containsAtLeast(u8, text, 1, "imported "));
+    try testing.expectEqualStrings("", captured.err.written());
+}
+
+// `runExport` needs a real data directory and the upstream probe leaves the
+// machine. Those cases are S7's:
+// `src/storage/storage_integration_test.zig` cases 20-22.

src/config/bootstrap.zig

@@ -7,7 +7,10 @@
 //! The policy is three lines long:
 //!
 //! - no file → normal steady state, keep the database as it is;
-//! - database already configured → the file is ignored, as PLAN §3.5 requires;
+//! - database already configured → the file is ignored, as PLAN §3.5 requires.
+//!   Configured means an operator put something there. A database that has only
+//!   answered queries is not configured, however many client rows the DNS path
+//!   materialised into it, and `import.isEmpty` is where that line is drawn;
 //! - otherwise → import it, and a file that is unreadable, unparseable or
 //!   invalid is an error. The operator wrote that file and meant it; starting
 //!   with silent defaults instead is the exact failure mode PLAN §1.3 exists to
@@ -55,6 +58,84 @@ pub fn bootstrap(
     return .seeded;
 }
 
-// Every path through `bootstrap` starts with a filesystem access, so all three
-// outcomes are exercised in `src/storage/storage_integration_test.zig` (S7)
-// against real files. There is nothing here that an in-memory test could reach.
+// ---------------------------------------------------------------------------
+// tests
+// ---------------------------------------------------------------------------
+//
+// All three outcomes are exercised end to end in
+// `src/storage/storage_integration_test.zig` (S7) against a real data directory.
+// What the two cases below add is the one distinction that decides which outcome
+// an operator gets, and it is too important to leave behind a `-Dintegration`
+// flag: whether the database has been *configured*, not whether it has been
+// *used*.
+
+const testing = std.testing;
+
+const clients_repo = @import("../storage/repositories/clients_repo.zig");
+const migrations = @import("../storage/migrations.zig");
+
+const seed_source =
+    \\.{
+    \\    .groups = .{ .{ .name = "default" } },
+    \\    .upstreams = .{ .{ .url = "https://dns.example/dns-query" } },
+    \\}
+;
+
+/// Unparseable on purpose: a call that succeeds proves the file was never read.
+const broken_source = ".{ .groups = ";
+
+fn openMigrated() !db.Db {
+    var database = try db.Db.open(":memory:", .{ .mode = .memory });
+    errdefer database.close();
+    try db.applyPragmas(&database, .{});
+    _ = try migrations.migrate(&database);
+    return database;
+}
+
+test "a server that has answered queries still seeds from its configuration file" {
+    const io = testing.io;
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    try tmp.dir.writeFile(io, .{ .sub_path = "config.zon", .data = seed_source });
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // The unattended first boot: the server came up on defaults, answered
+    // traffic, and the operator dropped a config file in afterwards.
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+
+    var diags: validate.Diagnostics = .init(testing.allocator);
+    defer diags.deinit();
+
+    const outcome = try bootstrap(io, testing.allocator, &database, tmp.dir, "config.zon", &diags);
+    try testing.expectEqual(Outcome.seeded, outcome);
+    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM upstreams"));
+    // Seeding did not cost the operator the device list they had been watching.
+    try testing.expectEqual(@as(i64, 1), try clients_repo.countClients(&database));
+}
+
+test "a client the operator has customised keeps the configuration file out" {
+    const io = testing.io;
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    try tmp.dir.writeFile(io, .{ .sub_path = "config.zon", .data = broken_source });
+
+    var database = try openMigrated();
+    defer database.close();
+
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+    const id = try database.queryInt("SELECT id FROM clients WHERE ip = '192.168.1.5'");
+    try clients_repo.updateClient(&database, id, .{ .name = "tv", .group_id = 1 });
+
+    var diags: validate.Diagnostics = .init(testing.allocator);
+    defer diags.deinit();
+
+    const outcome = try bootstrap(io, testing.allocator, &database, tmp.dir, "config.zon", &diags);
+    try testing.expectEqual(Outcome.db_already_configured, outcome);
+    try testing.expectEqual(@as(usize, 0), diags.problems.items.len);
+    // The name and the flag the operator set are still theirs.
+    try testing.expectEqual(@as(i64, 1), try database.queryInt(
+        "SELECT count(*) FROM clients WHERE name = 'tv' AND hand_edited = 1",
+    ));
+}

src/config/faults.zig

@@ -0,0 +1,115 @@
+//! One definition of "the operator's configuration is wrong".
+//!
+//! `run`, `check` and `import` all sort a failure into two buckets: the
+//! configuration is wrong and the operator can fix it (exit 2, `nxdns check`
+//! is the next step), or something else broke (exit 1). Each subcommand used to
+//! carry its own list of which errors meant which, and the lists disagreed —
+//! the same seed file exited 1 from `run` and 2 from `check`. There is one list
+//! now, and it is this file. Nothing else may keep a second one.
+
+const std = @import("std");
+
+const validate = @import("validate.zig");
+
+/// `ValidateError` enters as a whole set rather than variant by variant, so a
+/// variant added to the validator cannot silently fall through to exit 1. The
+/// four extras are the configuration faults raised outside the validator: the
+/// ZON reader (`ParseZon`), the seed-file size limit (`ConfigTooLarge`), the
+/// composition root's upstream build (`NoUsableUpstreams`) and its certificate
+/// load (`BadCertificate`).
+///
+/// Not here on purpose: `error.DatabaseNotEmpty`, which reports the state of
+/// the database rather than the content of a file, and is the one config-shaped
+/// exit 2 `cli` decides for itself.
+const ConfigFault = validate.ValidateError || error{
+    ParseZon,
+    ConfigTooLarge,
+    NoUsableUpstreams,
+    BadCertificate,
+};
+
+const faults: []const anyerror = blk: {
+    const set = @typeInfo(ConfigFault).error_set.?;
+    var list: [set.len]anyerror = undefined;
+    for (set, 0..) |member, i| list[i] = @field(anyerror, member.name);
+    const frozen = list;
+    break :blk &frozen;
+};
+
+/// True when `err` means the configuration the operator supplied is wrong.
+/// Linear over a set of about forty errors, on failure paths only.
+pub fn isConfigFault(err: anyerror) bool {
+    for (faults) |fault| {
+        if (err == fault) return true;
+    }
+    return false;
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+const testing = std.testing;
+
+test "every ValidateError variant is a configuration fault, with no exceptions" {
+    // The guard on the derivation: a variant added to `ValidateError` and left
+    // out of the classification fails here rather than exiting 1 in the field.
+    //
+    // No member is excused. This file used to subtract `error.OutOfMemory`
+    // here, which made the rule "every ValidateError is a fault, except one" —
+    // a private exclusion list of exactly the kind this file exists to abolish.
+    // `validate.ValidateError` no longer carries an allocation failure, so the
+    // rule is literal again.
+    inline for (@typeInfo(validate.ValidateError).error_set.?) |member| {
+        const err = @field(anyerror, member.name);
+        if (!isConfigFault(err)) {
+            std.debug.print("isConfigFault(error.{s}) is false, expected true\n", .{member.name});
+            return error.TestUnexpectedResult;
+        }
+    }
+}
+
+test "an allocation failure is not a member of the validator's verdict" {
+    // The root of it: `faults.zig` can only be exception-free while
+    // `ValidateError` holds nothing that is not a verdict on the file.
+    inline for (@typeInfo(validate.ValidateError).error_set.?) |member| {
+        if (std.mem.eql(u8, member.name, "OutOfMemory")) {
+            std.debug.print("ValidateError carries error.OutOfMemory\n", .{});
+            return error.TestUnexpectedResult;
+        }
+    }
+    // It is still reachable from `validate`, just not as a finding: the
+    // allocator can fail and the caller has to handle it.
+    comptime var reachable = false;
+    inline for (@typeInfo(validate.Error).error_set.?) |member| {
+        if (comptime std.mem.eql(u8, member.name, "OutOfMemory")) reachable = true;
+    }
+    try testing.expect(reachable);
+}
+
+test "the faults raised outside the validator are configuration faults" {
+    try testing.expect(isConfigFault(error.ParseZon));
+    try testing.expect(isConfigFault(error.ConfigTooLarge));
+    try testing.expect(isConfigFault(error.NoUsableUpstreams));
+    try testing.expect(isConfigFault(error.BadCertificate));
+}
+
+test "the seed-file errors that used to exit 1 from run are configuration faults" {
+    // D1 verbatim: these three reached `run` from a rejected seed file and were
+    // classified as runtime failures.
+    try testing.expect(isConfigFault(error.ParseZon));
+    try testing.expect(isConfigFault(error.MissingDefaultGroup));
+    try testing.expect(isConfigFault(error.NoUpstreams));
+}
+
+test "a runtime failure is not a configuration fault" {
+    try testing.expect(!isConfigFault(error.OutOfMemory));
+    try testing.expect(!isConfigFault(error.AccessDenied));
+    try testing.expect(!isConfigFault(error.FileNotFound));
+    try testing.expect(!isConfigFault(error.AddressInUse));
+    // A state conflict, not a bad file: `import` refuses to overwrite a
+    // configured database and decides that exit code itself.
+    try testing.expect(!isConfigFault(error.DatabaseNotEmpty));
+    // Only ever a warning, so it never reaches an exit code by this route.
+    try testing.expect(!isConfigFault(error.SourceInNoGroup));
+}

src/config/import.zig

@@ -1,4 +1,13 @@
-//! `nxdns import`: a ZON file becomes the whole content of `config.db`.
+//! `nxdns import`: a ZON file becomes the whole configuration of `config.db`.
+//!
+//! Configuration, not content: the `hand_edited = 0` client rows the DNS path
+//! materialises from live traffic are runtime state, they are absent from an
+//! export, and an import carries them across rather than deleting them.
+//!
+//! `clients.first_seen` and `clients.last_seen` are runtime state on every client
+//! row, configured ones included, and the config model carries neither. So an
+//! address the database already knew keeps both across an import, and only an
+//! address it has never seen takes the import's clock.
 //!
 //! The order is the specification. Nothing reaches the database until the file
 //! has been read, parsed and validated, and every write happens inside one
@@ -44,21 +53,31 @@ const canonical_buf_len = 64;
 // emptiness
 // ---------------------------------------------------------------------------
 
-/// A database is "never configured" when the migrations have run and nothing
-/// else has. The migrations themselves create `schema_version` and seed
-/// `groups(1, 'default')`, so "no rows anywhere" is the wrong test.
+/// A database is "never configured" when the migrations have run and the
+/// operator has added nothing. The migrations themselves create
+/// `schema_version` and seed `groups(1, 'default')`, so "no rows anywhere" is
+/// the wrong test.
 ///
-/// True iff every table in `config_schema.content_tables` is empty, `groups`
-/// holds exactly one row, and that row is the seeded `(1, 'default', 0)`.
+/// True iff no table in `config_schema.content_tables` holds a row the operator
+/// put there, `groups` holds exactly one row, and that row is the seeded
+/// `(1, 'default', 0)`.
 ///
-/// The client count here includes auto-materialised rows: a server that has
-/// answered one query is configured enough that a bootstrap file must not
-/// overwrite it.
+/// `clients` is the one table a row can reach without an operator: the DNS path
+/// materialises `hand_edited = 0` rows straight from live traffic (PLAN §7.2).
+/// Counting those made emptiness a function of traffic — a server that had
+/// answered a single query silently ignored the seed file its operator dropped
+/// next to it. So only `hand_edited = 1` rows count, which is the predicate
+/// `clients_repo.listClients` already exports by: this database is empty exactly
+/// when its export is empty.
 pub fn isEmpty(database: *db.Db) db.Error!bool {
     // `inline for` over a comptime table list: every statement below is a
     // compile-time string, so no table name is ever concatenated at run time.
     inline for (config_schema.content_tables) |table| {
-        if (try database.queryInt("SELECT count(*) FROM " ++ table) != 0) return false;
+        const count_sql = comptime if (std.mem.eql(u8, table, "clients"))
+            "SELECT count(*) FROM clients WHERE hand_edited = 1"
+        else
+            "SELECT count(*) FROM " ++ table;
+        if (try database.queryInt(count_sql) != 0) return false;
     }
     if (try database.queryInt("SELECT count(*) FROM groups") != 1) return false;
     const seeded = try database.queryInt(
@@ -179,6 +198,8 @@ pub fn applyToDb(
     // writes are one atomic unit.
     if (!options.force and !try isEmpty(database)) return error.DatabaseNotEmpty;
 
+    try liftSavedClients(database);
+
     inline for (config_schema.delete_order) |table| {
         try database.exec("DELETE FROM " ++ table ++ ";");
     }
@@ -203,6 +224,8 @@ pub fn applyToDb(
         canonical.ip = try canonicalIp(client.ip, &buf);
         try clients_repo.insertClient(database, canonical, ctx);
     }
+    try restoreSavedClients(database, group_ids.get("default").?);
+
     for (cfg.client_prefixes) |entry| {
         var buf: [canonical_buf_len]u8 = undefined;
         var canonical = entry;
@@ -238,6 +261,102 @@ pub fn applyToDb(
     try tx.commit();
 }
 
+const lift_saved_clients_sql: [:0]const u8 =
+    \\DROP TABLE IF EXISTS temp.saved_clients;
+    \\CREATE TEMP TABLE saved_clients AS
+    \\    SELECT ip, name, hand_edited, first_seen, last_seen FROM clients;
+;
+
+/// Carries what the wipe must not destroy over the wipe that follows: the
+/// auto-materialised client rows themselves, and the observed timestamps of
+/// every client row whatever its flag.
+///
+/// A `hand_edited = 0` row is runtime state, not configuration: the DNS path
+/// wrote it from live traffic and `clients_repo.listClients` already keeps it out
+/// of an export. Replacing the *configuration* must therefore not delete it —
+/// but it cannot survive in place either, because `clients.group_id` references
+/// `groups(id)` with no cascade, and the wipe empties `groups`. So the rows step
+/// aside into the temp database and come back once `groups` holds `default`
+/// again.
+///
+/// The table takes every row, not only the `hand_edited = 0` ones, because
+/// `first_seen` and `last_seen` are runtime state on a configured row too — the
+/// tracker keeps writing `last_seen` on the clients the operator named. Saving
+/// only the materialised rows would keep the observation history of the devices
+/// nobody named and destroy it for the devices somebody did. `hand_edited` rides
+/// along so the restore can tell the two apart.
+///
+/// `IF EXISTS` because a rolled-back import must not poison the next one.
+fn liftSavedClients(database: *db.Db) Error!void {
+    return database.exec(lift_saved_clients_sql);
+}
+
+const merge_observed_timestamps_sql: [:0]const u8 =
+    \\UPDATE clients AS c
+    \\   SET first_seen = m.first_seen, last_seen = m.last_seen
+    \\  FROM temp.saved_clients m
+    \\ WHERE m.ip = c.ip
+;
+
+const restore_materialised_clients_sql =
+    \\INSERT INTO clients (ip, name, group_id, hand_edited, first_seen, last_seen)
+    \\SELECT m.ip, m.name, ?1, 0, m.first_seen, m.last_seen
+    \\  FROM temp.saved_clients m
+    \\ WHERE m.hand_edited = 0
+    \\   AND NOT EXISTS (SELECT 1 FROM clients c WHERE c.ip = m.ip)
+;
+
+/// Puts back what `liftSavedClients` set aside, in two steps that must stay in
+/// this order: the restore drops the temp table on its way out, so the merge
+/// cannot follow it.
+///
+/// **The merge.** An address the config declares belongs to the config — but
+/// `first_seen` and `last_seen` are not the config's to state. The model carries
+/// neither field, so `insertClient` writes `now` into both as a placeholder for a
+/// device it knows nothing about. When the database already held that address, the
+/// placeholder is the worse of the two values and both columns come from the saved
+/// row instead. Everything else on the row stays the config's: the name, the
+/// group, and `hand_edited = 1`.
+///
+/// Not "the earlier `first_seen` and the later `last_seen`": the placeholder is
+/// the wall clock at import time and every real observation predates it, so
+/// "later" would resolve to the placeholder every time and stamp each named
+/// device as seen at the moment of the import. An operator restoring a backup
+/// would read that as liveness. `first_seen` and `last_seen` mean "when a query
+/// from this address arrived", `pruneStale` and the API both read them that way,
+/// and an import is not a query. Taking both from the saved row also keeps
+/// `first_seen <= last_seen`, which `upsertSeen` guarantees pairwise.
+///
+/// Only the config's own clients are in the table at this point, so the update
+/// needs no filter of its own, and the saved row's flag does not enter into it: a
+/// device keeps its history whether the previous row was materialised or
+/// configured. A `--force` re-import of the same file is the case that matters —
+/// it deletes and rewrites every configured client, and without the merge each
+/// one would come back claiming it was first seen at the moment of the import.
+///
+/// **The restore.** Only `hand_edited = 0` rows come back. A configured client is
+/// configuration, so a file that leaves its address out has removed that client
+/// and the row must stay gone; its history was saved for the merge, not for a
+/// resurrection. `WHERE NOT EXISTS` rather than `INSERT OR IGNORE`: the only
+/// materialised row worth skipping is one whose address the config claims — the
+/// operator naming a device the server had already discovered, handled by the
+/// merge above — and every other constraint failure stays loud.
+///
+/// The rows return to `default`, the group `upsertSeen` materialises into.
+/// `first_seen` and `last_seen` cross unchanged; the row id does not, because the
+/// config's clients went in first and hold the low ids now. Nothing references
+/// `clients.id` inside `config.db`, and §3.6 keeps the query log out of it.
+fn restoreSavedClients(database: *db.Db, default_group_id: i64) Error!void {
+    try database.exec(merge_observed_timestamps_sql);
+    {
+        var stmt = try database.prepare(restore_materialised_clients_sql);
+        defer stmt.deinit();
+        try stmt.bindInt(1, default_group_id);
+        try stmt.exec();
+    }
+    return database.exec("DROP TABLE temp.saved_clients;");
+}
+
 /// `default` goes in first and takes rowid 1. §11.2 seeds group 1 as `default`
 /// and §7.2's fallback assignment depends on it; letting an import renumber it
 /// would silently move every unassigned client.
@@ -303,7 +422,7 @@ fn canonicalPrefix(text: []const u8, buf: []u8) error{BadClientPrefix}![]const u
 }
 
 /// argon2id with the OWASP parameters (t=2, m=19 MiB, p=1) rather than the
-/// 64 MiB `interactive_2id`, because PLAN §18 budgets under 100 MB total on a
+/// 64 MiB `interactive_2id`, because PLAN §18 budgets under 100 MiB total on a
 /// Pi 5.
 ///
 /// `strHash`'s error set reaches beyond this module's (it carries
@@ -419,16 +538,45 @@ test "isEmpty is false once a settings row exists" {
     try testing.expect(!try isEmpty(&database));
 }
 
-test "isEmpty is false once an auto-materialized client exists" {
+test "isEmpty ignores the client rows live traffic materialises" {
     var database = try openMigrated();
     defer database.close();
-    try database.exec(
-        \\INSERT INTO clients (ip, name, group_id, hand_edited, first_seen, last_seen)
-        \\VALUES ('192.168.1.5', NULL, 1, 0, 1, 1);
+
+    // The real §7.2 write path, not hand-written SQL: what makes these rows
+    // ignorable is that `upsertSeen` is the thing that wrote them.
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+    try clients_repo.upsertSeen(&database, "192.168.1.6", 1700000100);
+
+    try testing.expectEqual(@as(i64, 2), try clients_repo.countClients(&database));
+    try testing.expect(try isEmpty(&database));
+}
+
+test "isEmpty is false once a client carries operator intent" {
+    var database = try openMigrated();
+    defer database.close();
+
+    _ = try clients_repo.insertClientRow(
+        &database,
+        .{ .ip = "192.168.1.7", .name = "printer", .group_id = 1 },
+        1700000000,
     );
     try testing.expect(!try isEmpty(&database));
 }
 
+test "isEmpty is false once an operator edits a materialised client" {
+    var database = try openMigrated();
+    defer database.close();
+
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+    try testing.expect(try isEmpty(&database));
+
+    // A PUT through the API is what turns a discovered device into policy, and
+    // that policy is exactly what a seed would replace.
+    const id = try database.queryInt("SELECT id FROM clients WHERE ip = '192.168.1.5'");
+    try clients_repo.updateClient(&database, id, .{ .name = "tv", .group_id = 1 });
+    try testing.expect(!try isEmpty(&database));
+}
+
 test "isEmpty is false once the seeded group is changed" {
     var database = try openMigrated();
     defer database.close();
@@ -459,6 +607,206 @@ test "importSource seeds a migrated database and group 'default' keeps id 1" {
     );
 }
 
+test "an import keeps the materialised clients it found and lets the config claim an address" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // Two devices the server discovered. `full_source` names the second one.
+    // The first is seen twice, so `first_seen` and `last_seen` differ and the
+    // assertions below cannot pass by carrying one column into both.
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000100);
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000200);
+
+    try importText(io, &database, full_source, .{});
+
+    try testing.expectEqual(@as(i64, 2), try clients_repo.countClients(&database));
+
+    // The device the config says nothing about keeps its flag, its group and
+    // both timestamps: a seed is not a reason to forget when a device appeared.
+    var stmt = try database.prepare(
+        "SELECT hand_edited, group_id, first_seen, last_seen FROM clients WHERE ip = '192.168.1.5'",
+    );
+    defer stmt.deinit();
+    try testing.expect(try stmt.step());
+    try testing.expectEqual(@as(i64, 0), stmt.columnInt(0));
+    try testing.expectEqual(@as(i64, 1), stmt.columnInt(1));
+    try testing.expectEqual(@as(i64, 1700000000), stmt.columnInt(2));
+    try testing.expectEqual(@as(i64, 1700000100), stmt.columnInt(3));
+
+    // The one the config names belongs to the config: named, in `kids`, and
+    // hand-edited, so a later prune leaves it alone.
+    try testing.expectEqual(@as(i64, 1), try database.queryInt(
+        \\SELECT count(*) FROM clients c JOIN groups g ON g.id = c.group_id
+        \\ WHERE c.ip = 'fd00::1' AND c.hand_edited = 1 AND c.name = 'tablet' AND g.name = 'kids'
+    ));
+}
+
+test "the config claiming a discovered address keeps that device's observed timestamps" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // The device is seen twice, so the two timestamps differ and neither
+    // assertion below can pass by carrying one column into the other.
+    // `full_source` names this address.
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000200);
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000900);
+
+    try importText(io, &database, full_source, .{});
+
+    var stmt = try database.prepare(
+        "SELECT hand_edited, name, first_seen, last_seen FROM clients WHERE ip = 'fd00::1'",
+    );
+    defer stmt.deinit();
+    try testing.expect(try stmt.step());
+    // The row is the config's: named, hand-edited.
+    try testing.expectEqual(@as(i64, 1), stmt.columnInt(0));
+    try testing.expectEqualStrings("tablet", stmt.columnText(1));
+    // The observation history is the tracker's. The import clock is `now`, so
+    // both columns would hold a value far above these if the import had written
+    // its own.
+    try testing.expectEqual(@as(i64, 1700000200), stmt.columnInt(2));
+    try testing.expectEqual(@as(i64, 1700000900), stmt.columnInt(3));
+}
+
+test "a failed import leaves the observed timestamps exactly as they were" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = testing.allocator;
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // One address the config below claims, one it says nothing about.
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000200);
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000900);
+    try clients_repo.upsertSeen(&database, "192.168.1.5", 1700000000);
+
+    const before = try dump(&database, gpa);
+    defer gpa.free(before);
+
+    // The clients go in, the timestamps merge, and then two identical local
+    // records violate `UNIQUE(name, rtype, value)`. Everything the import wrote
+    // must go with the transaction.
+    const broken: model.Config = .{
+        .groups = &.{.{ .name = "default" }},
+        .upstreams = &.{.{ .url = "https://dns.example/dns-query" }},
+        .clients = &.{.{ .ip = "fd00::1", .name = "tablet" }},
+        .local_records = &.{
+            .{ .name = "dup.lan", .rtype = .a, .value = "10.0.0.1" },
+            .{ .name = "dup.lan", .rtype = .a, .value = "10.0.0.1" },
+        },
+    };
+    try testing.expectError(error.Constraint, applyToDb(io, gpa, &database, broken, 42, .{}));
+
+    const after = try dump(&database, gpa);
+    defer gpa.free(after);
+    try testing.expectEqualStrings(before, after);
+}
+
+test "a forced re-import replaces the configured clients and keeps the materialised ones" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+
+    var database = try openMigrated();
+    defer database.close();
+    try importText(io, &database, full_source, .{});
+    try clients_repo.upsertSeen(&database, "10.0.0.9", 1700000200);
+    // The device the old file named keeps querying, so its row carries real
+    // observation history when the wipe reaches it.
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700003000);
+
+    try importText(io, &database, minimal_source, .{ .force = true });
+
+    // `full_source`'s hand-edited client went with the rest of the old
+    // configuration; the discovered one did not.
+    try testing.expectEqual(@as(i64, 1), try clients_repo.countClients(&database));
+    try testing.expectEqual(@as(i64, 1700000200), try database.queryInt(
+        "SELECT first_seen FROM clients WHERE ip = '10.0.0.9' AND hand_edited = 0",
+    ));
+    // The lift saves a configured client's history so the merge can hand it back
+    // to the same address. It must never become a reason to resurrect a client
+    // the new file leaves out: dropping a client from the file removes it.
+    try testing.expectEqual(@as(i64, 0), try database.queryInt(
+        "SELECT count(*) FROM clients WHERE ip = 'fd00::1'",
+    ));
+    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM upstreams"));
+}
+
+test "a forced re-import keeps the observed timestamps of a client the config names" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // The device appears in traffic, the operator's file names it, and it keeps
+    // querying afterwards. The row is now `hand_edited = 1` and carries a real
+    // `first_seen` and a later `last_seen`.
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700000200);
+    try importText(io, &database, full_source, .{});
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700005000);
+
+    try importText(io, &database, full_source, .{ .force = true });
+
+    var stmt = try database.prepare(
+        "SELECT hand_edited, first_seen, last_seen FROM clients WHERE ip = 'fd00::1'",
+    );
+    defer stmt.deinit();
+    try testing.expect(try stmt.step());
+    try testing.expectEqual(@as(i64, 1), stmt.columnInt(0));
+    // Re-importing the same file is not an observation of the device.
+    try testing.expectEqual(@as(i64, 1700000200), stmt.columnInt(1));
+    try testing.expectEqual(@as(i64, 1700005000), stmt.columnInt(2));
+}
+
+test "a failed forced import leaves every client timestamp exactly as it was" {
+    var threaded: std.Io.Threaded = .init(testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = testing.allocator;
+
+    var database = try openMigrated();
+    defer database.close();
+
+    // A configured client with observation history, and a materialised one.
+    try importText(io, &database, full_source, .{});
+    try clients_repo.upsertSeen(&database, "fd00::1", 1700005000);
+    try clients_repo.upsertSeen(&database, "10.0.0.9", 1700000200);
+
+    const before = try dump(&database, gpa);
+    defer gpa.free(before);
+
+    const broken: model.Config = .{
+        .groups = &.{.{ .name = "default" }},
+        .upstreams = &.{.{ .url = "https://dns.example/dns-query" }},
+        .clients = &.{.{ .ip = "fd00::1", .name = "tablet" }},
+        .local_records = &.{
+            .{ .name = "dup.lan", .rtype = .a, .value = "10.0.0.1" },
+            .{ .name = "dup.lan", .rtype = .a, .value = "10.0.0.1" },
+        },
+    };
+    try testing.expectError(
+        error.Constraint,
+        applyToDb(io, gpa, &database, broken, 42, .{ .force = true }),
+    );
+
+    const after = try dump(&database, gpa);
+    defer gpa.free(after);
+    try testing.expectEqualStrings(before, after);
+}
+
 test "applyToDb without force refuses a configured database and changes nothing" {
     var threaded: std.Io.Threaded = .init(testing.allocator, .{});
     defer threaded.deinit();

src/filter/filter_integration_test.zig

@@ -1041,17 +1041,78 @@ test "10: pruneOrphans deletes the files of a deleted source and leaves live one
     defer dir.close(io);
     try dir.writeFile(io, .{ .sub_path = "9999.list", .data = "gone.example.com\n" });
     try dir.writeFile(io, .{ .sub_path = "9999.wild", .data = "" });
-    // A refresh that is still running owns its temporaries, so they are not
-    // orphans and must survive.
+    // Id 9999 has no `blocklist_sources` row, so no refresh can be writing for
+    // it: these temporaries are what a refresh that died mid-write leaves
+    // behind, and the sweep is the only thing that will ever remove them.
     try dir.writeFile(io, .{ .sub_path = "9999.raw.tmp", .data = "" });
+    try dir.writeFile(io, .{ .sub_path = "9999.list.tmp", .data = "" });
+    try dir.writeFile(io, .{ .sub_path = "9999.wild.tmp", .data = "" });
+
+    // The live source does have a row, so its temporary is a refresh in
+    // progress and must survive a sweep that runs beside it.
+    var live_tmp_buf: [64]u8 = undefined;
+    const live_tmp = try std.fmt.bufPrint(&live_tmp_buf, "{d}.raw.tmp", .{id});
+    try dir.writeFile(io, .{ .sub_path = live_tmp, .data = "" });
 
     try env.mgr.pruneOrphans(io);
 
     var live_buf: [64]u8 = undefined;
     try dir.access(io, try std.fmt.bufPrint(&live_buf, "{d}.list", .{id}), .{});
-    try dir.access(io, "9999.raw.tmp", .{});
+    try dir.access(io, live_tmp, .{});
     try testing.expectError(error.FileNotFound, dir.access(io, "9999.list", .{}));
     try testing.expectError(error.FileNotFound, dir.access(io, "9999.wild", .{}));
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.raw.tmp", .{}));
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.list.tmp", .{}));
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.wild.tmp", .{}));
+}
+
+test "10b: the scheduler sweeps orphans on its own, with no operator call" {
+    if (!build_options.integration) return error.SkipZigTest;
+
+    const gpa = testing.allocator;
+    const env = try Env.create(gpa);
+    defer env.destroy();
+    const io = env.io();
+
+    const list_body = "aaa.example.com\n";
+    const id = try seedSource(&env.database, source_url);
+    try publishFixtureFiles(env, id, list_body, "");
+    // Dated now, so the startup pass finds nothing to download: this case
+    // asserts the sweep, and `source_url` points at a port nobody is serving.
+    try sources_repo.updateSourceStats(&env.database, id, .{
+        .last_updated = std.Io.Clock.real.now(io).toSeconds(),
+        .domain_count = 1,
+        .wildcard_count = 0,
+        .skipped_regex_count = 0,
+        .checksum = &bodyChecksum(list_body, ""),
+    });
+
+    var dir = try env.blocklistDir();
+    defer dir.close(io);
+    // What a source deleted while the server was down leaves, and what a
+    // process killed mid-refresh leaves. Nothing else in the tree removes
+    // either.
+    try dir.writeFile(io, .{ .sub_path = "9999.list", .data = "gone.example.com\n" });
+    try dir.writeFile(io, .{ .sub_path = "9999.wild", .data = "" });
+    try dir.writeFile(io, .{ .sub_path = "9999.raw.tmp", .data = "" });
+
+    // `runScheduler` is the entry point `app.zig` hands to `Io.Group`, and the
+    // only one the server ever calls. A disabled update stops it after the
+    // startup pass, so the production path runs to completion here with no
+    // interval to wait out.
+    env.mgr.update.enabled = false;
+    try env.mgr.runScheduler(io);
+
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.list", .{}));
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.wild", .{}));
+    try testing.expectError(error.FileNotFound, dir.access(io, "9999.raw.tmp", .{}));
+
+    // The live source kept its files and is still filtering: the sweep did not
+    // take the snapshot the same pass had just published.
+    var live_buf: [64]u8 = undefined;
+    try dir.access(io, try std.fmt.bufPrint(&live_buf, "{d}.list", .{id}), .{});
+    const decision, _ = try env.evaluate("aaa.example.com");
+    try testing.expect(decision.blocked);
 }
 
 // ---------------------------------------------------------------------------

src/filter/manager.zig

@@ -39,6 +39,7 @@ const std = @import("std");
 const Allocator = std.mem.Allocator;
 
 const model = @import("../config/model.zig");
+const safe_url = @import("../safe_url.zig");
 const db = @import("../storage/db.zig");
 const clients_repo = @import("../storage/repositories/clients_repo.zig");
 const groups_repo = @import("../storage/repositories/groups_repo.zig");
@@ -60,6 +61,10 @@ pub const max_error_len: usize = 128;
 /// `SourceStatus.url` is fixed-size so a copied status borrows nothing. A
 /// blocklist url longer than this is truncated in the status only; the row
 /// keeps it whole.
+///
+/// The log form of a url is bounded separately by `safe_url.max_len`. The two
+/// numbers agree today and answer different questions; neither follows the
+/// other.
 pub const max_url_len: usize = 255;
 
 /// A compiled body larger than this is refused at load. A source that reaches
@@ -80,6 +85,38 @@ const sample_buf_len: usize = parsers.sample_lines * (compiler.max_line_len + 1)
 /// `<id>` is at most 20 characters and the longest suffix is `.list.tmp`.
 const name_buf_len: usize = 48;
 
+/// How one blocklist source is named in a log line: by its row id and its name,
+/// which are its own identity, and by its redacted url, which says where it
+/// points and nothing more.
+///
+/// The url used to carry the identity here on its own. It cannot: `safe_url`
+/// drops the path, because a path segment is a place an operator's token lives,
+/// and two sources on one host are told apart by exactly that path. The id and
+/// the name are on the row every one of these lines already holds, they are
+/// what the API and the web UI show, and neither can leak what the url holds.
+/// The name is escaped for the same reason the url is — both are database text
+/// and a newline in either would forge a log line. It carries its own quotes,
+/// out of `safe_url.quoteText`, because a quote this format string added would
+/// be a quote the name could close: `ads' (https://decoy.example) --` would then
+/// read as a source pointing somewhere it does not.
+const SourceLabel = struct {
+    id: i64,
+    name: []const u8,
+    url: []const u8,
+
+    fn of(row: sources_repo.SourceRow) SourceLabel {
+        return .{ .id = row.id, .name = row.name, .url = row.url };
+    }
+
+    pub fn format(self: SourceLabel, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        try w.print("source {d} {f} {f}", .{
+            self.id,
+            safe_url.quoteText(self.name),
+            safe_url.redactQuoted(self.url),
+        });
+    }
+};
+
 pub const Paths = struct {
     /// `<data_dir>`, owned by the caller and left open for the manager's life.
     dir: std.Io.Dir,
@@ -520,7 +557,7 @@ pub const Manager = struct {
         // `replace` calls — a new `.list` beside an old `.wild` — is caught
         // here and refreshed, not served as a half-updated list.
         if (!std.mem.eql(u8, stored, &bodyChecksum(list_body, wild_body))) {
-            log.warn("blocklist {s}: compiled files do not match the stored checksum", .{row.url});
+            log.warn("blocklist {f}: compiled files do not match the stored checksum", .{SourceLabel.of(row)});
             return .{ .failed = .{ .state = .load_failed, .text = "ChecksumMismatch" } };
         }
 
@@ -610,7 +647,7 @@ pub const Manager = struct {
         defer self.deleteQuietly(io, dir, list_tmp);
         defer self.deleteQuietly(io, dir, wild_tmp);
 
-        self.download(io, dir, raw_name, row.url) catch |err| switch (err) {
+        self.download(io, dir, raw_name, row) catch |err| switch (err) {
             error.OutOfMemory => return error.OutOfMemory,
             error.Canceled => return error.Canceled,
             else => {
@@ -692,13 +729,16 @@ pub const Manager = struct {
     }
 
     /// The body goes to a temporary file, never to memory: `max_body_bytes` is
-    /// 64 MB and the memory budget has no room for it beside two snapshots.
+    /// 64 MiB and the memory budget has no room for it beside two snapshots.
+    ///
+    /// It takes the whole row rather than the url alone because its two log
+    /// lines name the source by its id and name, which only the row carries.
     fn download(
         self: *Manager,
         io: std.Io,
         dir: std.Io.Dir,
         raw_name: []const u8,
-        url: []const u8,
+        row: sources_repo.SourceRow,
     ) !void {
         const file = try dir.createFile(io, raw_name, .{ .permissions = .fromMode(0o600) });
         defer file.close(io);
@@ -707,14 +747,14 @@ pub const Manager = struct {
         defer self.gpa.free(buffer);
 
         var fw = file.writer(io, buffer);
-        const result = self.fetchWithin(io, url, &fw.interface) catch |err| {
+        const result = self.fetchWithin(io, row.url, &fw.interface) catch |err| {
             // `fetcher.Error.Unexpected` is what a failing sink surfaces as;
             // the concrete cause is on this writer, which the fetcher does not
             // own.
             if (fw.err) |cause| return cause;
             if (err == error.HttpStatus) {
                 if (self.fetcher.last_status) |status| {
-                    log.warn("blocklist {s}: http status {d}", .{ url, @intFromEnum(status) });
+                    log.warn("blocklist {f}: http status {d}", .{ SourceLabel.of(row), @intFromEnum(status) });
                 }
             }
             return err;
@@ -724,7 +764,7 @@ pub const Manager = struct {
         // buffer this function is about to drop.
         try file.sync(io);
 
-        log.debug("blocklist {s}: downloaded {d} bytes", .{ url, result.bytes_read });
+        log.debug("blocklist {f}: downloaded {d} bytes", .{ SourceLabel.of(row), result.bytes_read });
     }
 
     /// `std.http.Client` has no per-request deadline, so the whole exchange
@@ -906,7 +946,7 @@ pub const Manager = struct {
         err: anyerror,
     ) void {
         _ = self;
-        log.warn("blocklist {s}: download failed: {s}", .{ row.url, @errorName(err) });
+        log.warn("blocklist {f}: download failed: {s}", .{ SourceLabel.of(row), @errorName(err) });
         status.fail(.fetch_failed, @errorName(err));
     }
 
@@ -917,7 +957,7 @@ pub const Manager = struct {
         err: anyerror,
     ) void {
         _ = self;
-        log.warn("blocklist {s}: compile failed: {s}", .{ row.url, @errorName(err) });
+        log.warn("blocklist {f}: compile failed: {s}", .{ SourceLabel.of(row), @errorName(err) });
         status.fail(.compile_failed, @errorName(err));
     }
 
@@ -934,7 +974,7 @@ pub const Manager = struct {
             "NoValidEntries invalid={d} unsupported={d} long_lines={d}",
             .{ counts.invalid, counts.skipped_unsupported, counts.long_lines },
         ) catch "NoValidEntries";
-        log.warn("blocklist {s}: {s}", .{ row.url, text });
+        log.warn("blocklist {f}: {s}", .{ SourceLabel.of(row), text });
         status.fail(.no_valid_entries, text);
     }
 
@@ -954,6 +994,13 @@ pub const Manager = struct {
     /// `update.enabled == false` stops after the startup pass; manual refresh
     /// through `refreshAll` still works.
     pub fn runScheduler(self: *Manager, io: std.Io) std.Io.Cancelable!void {
+        // Ahead of the pass, not after it. This is the sweep that collects what
+        // a killed process left behind: a `.raw.tmp` as large as the body the
+        // dead refresh was writing, and the compiled files of a source deleted
+        // while the server was down. Both are bytes the pass below is about to
+        // ask the same filesystem for.
+        try self.sweepOrphans(io);
+
         self.startupPass(io) catch |err| switch (err) {
             error.Canceled => return error.Canceled,
             else => log.warn("blocklist startup pass failed: {s}", .{@errorName(err)}),
@@ -968,6 +1015,11 @@ pub const Manager = struct {
         };
         while (true) {
             try interval.sleep(io);
+            // Ahead of the gate as well as ahead of the pass: the sweep only
+            // unlinks, so it is the one thing here that can give a critically
+            // full disk room back, and gating it would keep the residue that
+            // helped fill the disk in the first place.
+            try self.sweepOrphans(io);
             if (self.refreshGated()) continue;
             self.refreshAll(io) catch |err| switch (err) {
                 error.Canceled => return error.Canceled,
@@ -976,6 +1028,21 @@ pub const Manager = struct {
         }
     }
 
+    /// `pruneOrphans` with its failure absorbed. Leftover bytes under
+    /// `<data_dir>/blocklists/` are not an outage, and a sweep that could not
+    /// read the directory must not cost the household the refresh pass behind
+    /// it — let alone the server. Cancellation is the one outcome that
+    /// propagates, because it means shutdown.
+    ///
+    /// Taken from outside every `*Locked` body: `pruneOrphans` takes
+    /// `writer_lock` itself and the mutex is not reentrant.
+    fn sweepOrphans(self: *Manager, io: std.Io) std.Io.Cancelable!void {
+        self.pruneOrphans(io) catch |err| switch (err) {
+            error.Canceled => return error.Canceled,
+            else => log.warn("pruning orphaned blocklist files failed: {s}", .{@errorName(err)}),
+        };
+    }
+
     /// The §11.6 gate, consulted by scheduled passes only (ruling 17). A
     /// download writes tens of megabytes into the blocklist directory and the
     /// compile writes as much again, which is exactly the "non-essential write"
@@ -1045,12 +1112,26 @@ pub const Manager = struct {
     // orphans
     // -----------------------------------------------------------------------
 
-    /// Deletes `<id>.list` and `<id>.wild` files whose id is no longer a
-    /// `blocklist_sources` row. Files of a live source are left alone,
+    /// Deletes the compiled files and the leftover temporaries whose id is no
+    /// longer a `blocklist_sources` row. Files of a live source are left alone,
     /// whatever their state.
+    ///
+    /// Three callers, and between them they cover every way an orphan is made:
+    /// `runScheduler` sweeps once before its startup pass — the residue of a
+    /// process that was killed mid-refresh, and of a source deleted while the
+    /// server was down — and again before each scheduled pass; the
+    /// `DELETE /api/blocklists/{id}` handler sweeps as soon as it has removed
+    /// the row, so the directory follows the table an operator can see instead
+    /// of waiting out `blocklist_update.interval_hours`.
+    ///
+    /// It is safe to call on a fresh install: `openDir` creates
+    /// `<data_dir>/blocklists/` if nothing has yet, and an empty directory
+    /// sweeps to nothing.
     pub fn pruneOrphans(self: *Manager, io: std.Io) Error!void {
-        // A refresh in flight owns the temporaries of a live source; the sweep
-        // must not run beside one and decide from a half-written directory.
+        // Every path that writes a temporary holds this lock too, so the sweep
+        // never reads a directory a refresh is halfway through. The temporaries
+        // it can see therefore belong to a finished or a dead refresh, and only
+        // those of a source with no row are removed.
         self.writer_lock.lockUncancelable(io);
         defer self.writer_lock.unlock(io);
 
@@ -1079,14 +1160,14 @@ pub const Manager = struct {
                 },
             } orelse break;
             if (entry.kind != .file) continue;
-            const id = compiledId(entry.name) orelse continue;
+            const id = sourceFileId(entry.name) orelse continue;
             if (containsId(rows.items, id)) continue;
             try doomed.append(self.gpa, try self.gpa.dupe(u8, entry.name));
         }
 
         for (doomed.items) |name| {
             self.deleteQuietly(io, dir, name);
-            log.info("pruned orphaned compiled file {s}", .{name});
+            log.info("pruned orphaned blocklist file {s}", .{name});
         }
     }
 
@@ -1253,7 +1334,7 @@ const LoadOutcome = union(enum) {
 };
 
 fn loadFailure(row: sources_repo.SourceRow, file_name: []const u8, err: anyerror) LoadOutcome {
-    log.warn("blocklist {s}: reading {s} failed: {s}", .{ row.url, file_name, @errorName(err) });
+    log.warn("blocklist {f}: reading {s} failed: {s}", .{ SourceLabel.of(row), file_name, @errorName(err) });
     return .{ .failed = .{ .state = .load_failed, .text = @errorName(err) } };
 }
 
@@ -1394,17 +1475,27 @@ fn compiledName(buf: *[name_buf_len]u8, id: i64, suffix: []const u8) []const u8
     return std.fmt.bufPrint(buf, "{d}{s}", .{ id, suffix }) catch unreachable;
 }
 
-/// The source id a compiled file belongs to, or null when the name is not one
-/// of ours. Temporary files are deliberately not matched: they belong to a
-/// refresh that may still be running.
-fn compiledId(file_name: []const u8) ?i64 {
-    const stem = if (std.mem.endsWith(u8, file_name, ".list"))
-        file_name[0 .. file_name.len - ".list".len]
-    else if (std.mem.endsWith(u8, file_name, ".wild"))
-        file_name[0 .. file_name.len - ".wild".len]
-    else
-        return null;
-    return std.fmt.parseInt(i64, stem, 10) catch null;
+/// Every name `compiledName` can produce, longest suffix first so `.list.tmp`
+/// is never read as `.list`.
+const source_file_suffixes = [_][]const u8{ ".list.tmp", ".wild.tmp", ".raw.tmp", ".list", ".wild" };
+
+/// The source id a file under the blocklist directory belongs to, or null when
+/// the name is not one of ours.
+///
+/// The three temporaries count. A refresh that dies between writing one and
+/// renaming it leaves a file no later refresh reuses and no `defer` reaches, so
+/// excluding them from the sweep means nothing ever removes them. Matching them
+/// is safe because `pruneOrphans` holds `writer_lock` for its whole body: every
+/// path that creates a temporary runs under that same lock, so no refresh is in
+/// flight while the sweep reads the directory, and a temporary the sweep sees
+/// belonging to a source that still has a row is kept regardless.
+fn sourceFileId(file_name: []const u8) ?i64 {
+    for (source_file_suffixes) |suffix| {
+        if (!std.mem.endsWith(u8, file_name, suffix)) continue;
+        const stem = file_name[0 .. file_name.len - suffix.len];
+        return std.fmt.parseInt(i64, stem, 10) catch null;
+    }
+    return null;
 }
 
 fn containsId(rows: []const sources_repo.SourceRow, id: i64) bool {
@@ -1566,6 +1657,62 @@ test "the header writer produces the documented text" {
     ++ "# sha256 " ++ "0" ** 64 ++ "\n", w.buffered());
 }
 
+test "the log label names a source without printing what its url carries" {
+    // Every `log.warn` in this file formats its subject through `SourceLabel`,
+    // so this is the text of those lines. A `std.log` line is not observable
+    // from a unit test under the default runner; the label is.
+    var buf: [1024]u8 = undefined;
+
+    const row: sources_repo.SourceRow = .{
+        .id = 3,
+        .url = "https://lists.example/download/token/hunter2/hosts.txt?apikey=s3cr3t",
+        .name = "ads",
+        .enabled = true,
+        .last_updated = null,
+        .domain_count = 0,
+        .wildcard_count = 0,
+        .skipped_regex_count = 0,
+        .checksum = null,
+    };
+    const printed = try std.fmt.bufPrint(&buf, "blocklist {f}: download failed: {s}", .{
+        SourceLabel.of(row),
+        @errorName(error.ConnectFailed),
+    });
+    try testing.expectEqualStrings(
+        "blocklist source 3 'ads' 'https://lists.example': download failed: ConnectFailed",
+        printed,
+    );
+    try testing.expect(!std.mem.containsAtLeast(u8, printed, 1, "hunter2"));
+    try testing.expect(!std.mem.containsAtLeast(u8, printed, 1, "s3cr3t"));
+
+    // The row is database text, and a path that writes it does not have to
+    // validate as strictly as the config validator. Neither column may end the
+    // line and start one of the operator's choosing.
+    var forged = row;
+    forged.name = "ads\n2026-01-01 ERROR forged";
+    forged.url = "https://lists.example\n2026-01-01 ERROR forged/hosts.txt";
+    const escaped = try std.fmt.bufPrint(&buf, "blocklist {f}", .{SourceLabel.of(forged)});
+    try testing.expectEqualStrings(
+        "blocklist source 3 'ads\\n2026-01-01 ERROR forged'" ++
+            " 'https://lists.example\\n2026-01-01 ERROR forged'",
+        escaped,
+    );
+    try testing.expect(!std.mem.containsAtLeast(u8, escaped, 1, "\n"));
+
+    // A name is operator-supplied and reaches the row through the API, so it
+    // can close the quote this label puts around it and open a decoy that reads
+    // as the url of a second source. The quote it would close is escaped, and
+    // the escape is unambiguous because a `\` is escaped too.
+    var decoy = row;
+    decoy.name = "ads' (https://decoy.example) --";
+    decoy.url = "https://lists.example/hosts.txt";
+    const quoted = try std.fmt.bufPrint(&buf, "blocklist {f}", .{SourceLabel.of(decoy)});
+    try testing.expectEqualStrings(
+        "blocklist source 3 'ads\\' (https://decoy.example) --' 'https://lists.example'",
+        quoted,
+    );
+}
+
 test "stripHeader returns the body of a compiled file" {
     const file =
         "# nxdns blocklist\n" ++
@@ -1615,13 +1762,28 @@ test "compiledName spells the four file names of a source" {
     try testing.expectEqualStrings("42.list.tmp", compiledName(&buf, 42, ".list.tmp"));
 }
 
-test "compiledId matches compiled files and nothing else" {
-    try testing.expectEqual(@as(?i64, 7), compiledId("7.list"));
-    try testing.expectEqual(@as(?i64, 7), compiledId("7.wild"));
-    try testing.expectEqual(@as(?i64, null), compiledId("7.list.tmp"));
-    try testing.expectEqual(@as(?i64, null), compiledId("7.raw.tmp"));
-    try testing.expectEqual(@as(?i64, null), compiledId("notes.list"));
-    try testing.expectEqual(@as(?i64, null), compiledId("README"));
+test "sourceFileId matches every name a refresh writes, including the temporaries" {
+    try testing.expectEqual(@as(?i64, 7), sourceFileId("7.list"));
+    try testing.expectEqual(@as(?i64, 7), sourceFileId("7.wild"));
+
+    // A temporary left by a refresh that died belongs to its source id, so the
+    // sweep can tell whether that source still has a row.
+    try testing.expectEqual(@as(?i64, 7), sourceFileId("7.raw.tmp"));
+    try testing.expectEqual(@as(?i64, 7), sourceFileId("7.list.tmp"));
+    try testing.expectEqual(@as(?i64, 7), sourceFileId("7.wild.tmp"));
+
+    try testing.expectEqual(@as(?i64, null), sourceFileId("notes.list"));
+    try testing.expectEqual(@as(?i64, null), sourceFileId("notes.raw.tmp"));
+    try testing.expectEqual(@as(?i64, null), sourceFileId("7.tmp"));
+    try testing.expectEqual(@as(?i64, null), sourceFileId("7.raw"));
+    try testing.expectEqual(@as(?i64, null), sourceFileId("README"));
+}
+
+test "every name compiledName writes is a name the sweep can attribute" {
+    var buf: [name_buf_len]u8 = undefined;
+    for (source_file_suffixes) |suffix| {
+        try testing.expectEqual(@as(?i64, 42), sourceFileId(compiledName(&buf, 42, suffix)));
+    }
 }
 
 test "a failed refresh keeps the fields of the compiled files still serving" {

src/main.zig

@@ -18,13 +18,35 @@ comptime {
     _ = @import("cli.zig");
 }
 
+/// Streaming, never positional, and this is load-bearing rather than a default
+/// worth changing back.
+///
+/// `File.writer` builds a `.positional` writer: it keeps an offset of its own,
+/// starting at zero, and pwrites there. `std.log` writes to the same descriptor
+/// directly and shares the descriptor's offset. With stderr redirected to a
+/// regular file the two disagree about where the end is, so the runner's first
+/// flush lands on top of whatever the log already wrote at the start and the
+/// operator loses those lines. `writerStreaming` uses the descriptor's own
+/// offset, so both writers append.
+///
+/// A terminal and a pipe are unaffected either way, because neither is seekable
+/// and the positional writer falls back to streaming on them. `2>file` is where
+/// it shows, and `nxdns run 2>logfile` is an ordinary way to run this program.
+/// `>>` needs it too: pwrite ignores `O_APPEND`.
+///
+/// Takes the `File` rather than naming stdout and stderr inside, so a test can
+/// hand it a real file and read back what an operator would have.
+fn runnerWriter(file: std.Io.File, io: std.Io, buffer: []u8) std.Io.File.Writer {
+    return file.writerStreaming(io, buffer);
+}
+
 pub fn main(init: std.process.Init) u8 {
     // Both `File.Writer` values are self-referential and must not move, so they
     // stay in these `var` slots for the whole of `main`.
     var out_buffer: [4096]u8 = undefined;
     var err_buffer: [4096]u8 = undefined;
-    var out = std.Io.File.stdout().writer(init.io, &out_buffer);
-    var err = std.Io.File.stderr().writer(init.io, &err_buffer);
+    var out = runnerWriter(std.Io.File.stdout(), init.io, &out_buffer);
+    var err = runnerWriter(std.Io.File.stderr(), init.io, &err_buffer);
 
     const runner: cli.Runner = .{
         .io = init.io,
@@ -55,3 +77,39 @@ pub fn main(init: std.process.Init) u8 {
         .help => cli.runHelp(runner),
     };
 }
+
+// ---------------------------------------------------------------------------
+// tests
+// ---------------------------------------------------------------------------
+
+test "a runner writer appends after what std.log already put in a redirected file" {
+    // Reproduced live before this test existed: `nxdns run --config <bad file>
+    // 2> file` lost `info(migrations): config.db migrated from schema version 0
+    // to 2`, because the runner's error writer pwrote its first flush at offset
+    // zero, over the line the log had already written. Down a pipe the same run
+    // kept both lines, which is why a terminal never showed this.
+    var threaded: std.Io.Threaded = .init(std.testing.allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    const gpa = std.testing.allocator;
+
+    var tmp = std.testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var file = try tmp.dir.createFile(io, "stderr.txt", .{});
+    defer file.close(io);
+
+    // What the log sink does: a direct write on the shared descriptor, which
+    // moves the descriptor's offset. A positional writer does not see it.
+    const log_line = "info(migrations): config.db migrated from schema version 0 to 2\n";
+    try file.writeStreamingAll(io, log_line);
+
+    const runner_line = "nxdns run failed: MissingDefaultGroup\n";
+    var buffer: [4096]u8 = undefined;
+    var w = runnerWriter(file, io, &buffer);
+    try w.interface.writeAll(runner_line);
+    try w.interface.flush();
+
+    const seen = try tmp.dir.readFileAlloc(io, "stderr.txt", gpa, .limited(4096));
+    defer gpa.free(seen);
+    try std.testing.expectEqualStrings(log_line ++ runner_line, seen);
+}

src/safe_url.zig

@@ -0,0 +1,947 @@
+//! The one redaction an operator-supplied string passes through before it
+//! reaches a log line or an operator-facing diagnostic.
+//!
+//! It lives at the root rather than under `filter/` because both the blocklist
+//! manager and the configuration validator print urls, and a `config/` module
+//! importing `filter/` would be the wrong dependency direction. It takes bytes
+//! and returns bytes: no `Io`, no allocator, no failure mode.
+
+const std = @import("std");
+
+/// The longest text one `SafeUrl` or one `QuotedText` prints, ahead of the `...`
+/// that marks a truncation. An operator-supplied url or name has no length
+/// limit and a log line must have one. A `QuotedText`'s two quotes fall outside
+/// the count: they are the delimiter rather than the text, and a truncated name
+/// still closes the one it opened.
+///
+/// The count is of printed characters, not of source bytes: an escape sequence
+/// costs the two or four characters it prints, not the one byte it stands for.
+/// A string of control characters therefore truncates early instead of printing
+/// several times its length.
+///
+/// This is deliberately a second number rather than a reuse of the blocklist
+/// manager's `max_url_len`, which sizes the inline url copy inside
+/// `SourceStatus`. The two answer different questions — how wide one log line
+/// may be, and how large a status value a caller may keep is — and they agree
+/// on 255 only because that width suits both. Neither may be changed on the
+/// other's account.
+pub const max_len: usize = 255;
+
+/// The only form of a url that may be written to a log. `format` prints the
+/// scheme, the host and the port; it drops the userinfo, the path, the query
+/// and the fragment; it escapes every control character; and it bounds the
+/// result at `max_len`.
+///
+/// Where the line falls, and why it falls there. A blocklist source url is
+/// operator-supplied and nothing on the way in restricts it to a bare public
+/// path: `?apikey=…`, a signed url whose signature is a query parameter,
+/// `https://user:pass@host/list.txt` and `https://host/d/<token>/hosts.txt` all
+/// validate, import and store. Userinfo, path, query and fragment are the four
+/// components where a credential can legally live, and a log line outlives the
+/// process — journald and the configured log file keep it, and neither is as
+/// protected as the database row the url came from. The rule against writing a
+/// secret to a log is absolute, so all four go.
+///
+/// Dropping the path costs the url the one job it used to do here: telling two
+/// sources on one host apart. That job was never the url's. A caller names the
+/// source it reports on from the source's own identity — `filter/manager.zig`
+/// prints the row id and the name beside this — and an identity read from the
+/// row cannot leak what the url holds.
+///
+/// What it still prints, deliberately: the scheme, the host and the port,
+/// because an operator reading a failure has to know where the source points;
+/// and, for a url that names no scheme at all, whatever precedes the first `/`
+/// once the userinfo is removed, because nothing in such a string distinguishes
+/// a host from anything else and that text is the closest thing to one. A
+/// string that names a scheme without a `//` after it gets no such treatment,
+/// whether or not it holds an `@` — see `redact`, where the text after such a
+/// scheme is a path segment under at least one reading and is never printed.
+///
+/// The host is therefore the one place a credential still survives this, and it
+/// is not hypothetical: NextDNS identifies an account by a profile id, which its
+/// DoH url carries in the path — `https://dns.nextdns.io/abcd12`, dropped here —
+/// and its DoT url carries in the hostname — `tls://abcd12.dns.nextdns.io`,
+/// kept. Nothing can be done about the second without printing no host at all,
+/// which would leave every line unactionable. An operator whose provider puts a
+/// secret in the hostname has published it to every resolver on the way to it
+/// long before it reaches this function.
+///
+/// It scans and never parses, so it cannot fail. `error.BadUrl` — the error
+/// `std.Uri.parse` raises — is one of the failures reported through here, so
+/// the inputs a parser refuses are exactly the inputs this must still redact.
+/// Only `://` introduces an authority, though: a run of one separator, of three
+/// or more, or of two that are not both `/`, leaves text that RFC 3986 reads as a
+/// path and WHATWG may read as a host, and `redact` prints no host it cannot
+/// settle. A `\` still ends an authority and still anchors the scheme scan — see
+/// `isSeparator` — it just does not open one.
+///
+/// Scanning a string a parser refuses means some of those strings have more
+/// than one reading, and the readings disagree about which side of an `@` the
+/// host is on. `https://lists.example?token=prefix@hunter2` is one: RFC 3986
+/// ends the authority at the `?`, so the host is `lists.example` and `hunter2`
+/// is part of an api key; read the `@` as a userinfo delimiter instead and the
+/// host is `hunter2`. No scan resolves that, and for a while this one chose the
+/// second reading and printed `https://hunter2` — a query-string secret printed
+/// as a host, which is the exact leak this file exists to stop. `authority` is
+/// therefore `null` on such a url and `format` prints
+/// `(ambiguous authority omitted)` in place of a host. Withholding the
+/// authority is always available and never wrong; choosing a side is wrong half
+/// the time, and the half it is wrong in is the half that leaks.
+pub const SafeUrl = struct {
+    /// The scheme without its `:`, or empty when the url carries no scheme
+    /// delimiter.
+    scheme: []const u8,
+    /// The authority without its userinfo — the host and, when present, the
+    /// port — or `null` when the url admits two readings of where the host is.
+    /// Empty and `null` are different answers: empty says the url names no
+    /// authority, `null` says it names one this cannot resolve.
+    authority: ?[]const u8,
+
+    /// Unquoted. A caller that prints this inside a delimiter of its own must
+    /// use `redactQuoted` instead, or handle the delimiter itself the way
+    /// `web/metrics.zig` does for a Prometheus label value.
+    ///
+    /// This renders `scheme://authority` canonically; it does not quote the
+    /// input's own syntax. A scheme is always followed by `://`, so
+    /// `tls:\\host` prints `tls://` and `https://?apikey=…` prints `https://`,
+    /// though only the second contained a `//`. The `://` marks where a host
+    /// would go, and the operator's remedy is the same either way: the url names
+    /// no host this can print. Nothing about which bytes separated the scheme
+    /// from the rest survives redaction, and nothing should — the input is not
+    /// reproducible from this and is not meant to be.
+    pub fn format(self: SafeUrl, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        return self.write(w, .none);
+    }
+
+    /// A truncation ends the value and returns, so a caller that has written an
+    /// opening delimiter still gets to write the closing one.
+    fn write(self: SafeUrl, w: *std.Io.Writer, delimiter: Delimiter) std.Io.Writer.Error!void {
+        var budget: usize = max_len;
+        if (!try writeEscaped(w, self.scheme, &budget, delimiter)) return w.writeAll("...");
+        if (self.scheme.len != 0 and !try writeEscaped(w, "://", &budget, delimiter))
+            return w.writeAll("...");
+        const authority = self.authority orelse ambiguous_authority;
+        if (!try writeEscaped(w, authority, &budget, delimiter)) return w.writeAll("...");
+    }
+};
+
+/// A redacted url in the form a caller may print inside a log line or a
+/// diagnostic sentence, where it needs a delimiter to keep a host from running
+/// into the words around it.
+///
+/// **`format` writes the quotes**, for the reason `QuotedText` states: a
+/// delimiter a caller adds is a delimiter the value can close. Redaction does
+/// not make that go away. A `'` is neither a component separator nor a control
+/// character, so it survives the scan into the authority — `https://ho'st/x`
+/// redacts to `https://ho'st`, and inside a caller's own `'…'` that reads as
+/// `'ho'` followed by loose text. The quote and its escape therefore live here,
+/// together, and a caller adds none of its own.
+///
+/// A caller sizing a fixed buffer needs `max_len + 3` for the value, as
+/// `SafeUrl` does, plus the two quotes.
+pub const QuotedUrl = struct {
+    url: SafeUrl,
+
+    pub fn format(self: QuotedUrl, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        try w.writeByte('\'');
+        try self.url.write(w, .single_quote);
+        try w.writeByte('\'');
+    }
+};
+
+/// What `format` prints where a host would go when it cannot say which text is
+/// the host. It is prose rather than a placeholder host because an operator has
+/// to read it as a statement about the line and not as an address: `https://`
+/// alone already means "this url names no authority", and the two call for
+/// different actions.
+///
+/// It spends the same printing budget the authority would have, so the bound
+/// `web/metrics.zig` sizes its buffer against still holds.
+///
+/// An operator can write a source url whose redaction is this same text, since
+/// a url that is not a url prints as itself. That collision costs nothing: it
+/// makes one line say less about a source than it could, and it cannot make a
+/// credential read as a host, which is the direction that matters.
+const ambiguous_authority = "(ambiguous authority omitted)";
+
+/// An operator-supplied string that is not a url — a blocklist source name — in
+/// the only form it may be written to a log. It holds no credential by design,
+/// so nothing is dropped from it; it comes out of a database row the same way a
+/// url does, so a control character in it can forge a log line the same way,
+/// and `format` escapes and bounds it for that reason alone.
+///
+/// **`format` writes the quotes.** A caller printing a name inside a log line
+/// has to delimit it, or a name with a space in it runs into the words around
+/// it; and a delimiter a caller adds is a delimiter the name can close. `ads'
+/// (https://decoy.example) --` inside a caller's quotes produces a line naming a
+/// url no source has. So the quote and its escape live in one place, here,
+/// where they cannot drift apart. A caller adds none of its own.
+pub const QuotedText = struct {
+    text: []const u8,
+
+    pub fn format(self: QuotedText, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        var budget: usize = max_len;
+        try w.writeByte('\'');
+        if (!try writeEscaped(w, self.text, &budget, .single_quote)) try w.writeAll("...");
+        try w.writeByte('\'');
+    }
+};
+
+/// The `SafeUrl` of `url`. Both fields borrow from `url`, which every caller
+/// holds for the length of the call it prints in.
+pub fn redact(url: []const u8) SafeUrl {
+    const first_sep = std.mem.indexOfAny(u8, url, separators) orelse url.len;
+    const first_colon = std.mem.indexOfScalar(u8, url, ':') orelse url.len;
+
+    // A scheme delimiter is a colon immediately before the first separator of
+    // the whole string. Anchoring on the first separator is what keeps `a/b:/c`
+    // from reading as a scheme. It anchors on a `\` too, so a url pasted with
+    // backslashes still has its scheme recognised and echoed, even though a `\`
+    // no longer opens an authority.
+    var scheme: []const u8 = "";
+    var rest = url;
+    var delimited = false;
+    if (first_colon + 1 == first_sep and isScheme(url[0..first_colon])) {
+        scheme = url[0..first_colon];
+        var after = first_sep;
+        while (after < url.len and isSeparator(url[after])) after += 1;
+        // Exactly `//` introduces an authority, and nothing else does. One
+        // separator leaves an absolute path — RFC 3986 reads `https:/hunter2` as
+        // the path `/hunter2`, WHATWG reads `hunter2` as the host — and three or
+        // more is an empty authority to the first and a host to the second.
+        //
+        // A run of two that is not two slashes depends on the scheme. WHATWG
+        // converts a `\` to a `/` only for a *special* scheme, so `https:\\host`
+        // is contested the same way, while `tls:\\host` — `tls` is not special —
+        // has no reading at all under which `host` is a host. RFC 3986 gives `\`
+        // no meaning anywhere.
+        //
+        // So the two answers are different answers, and `authority` carries the
+        // difference: `null` where the readings disagree, empty where they agree
+        // the url names no authority. Either way the path segment stays out of
+        // the log, which is the property that matters; this decides only what the
+        // line then claims about it.
+        //
+        // An earlier revision accepted any number of separators, to keep
+        // `https:/user:pass@host/list` from printing its userinfo. That reason
+        // expired when the `/` cut moved ahead of the userinfo lookup: the
+        // authority of a url with no `://` now ends at its first `/`, so it holds
+        // no userinfo to print. Withholding it is both safe and the honest
+        // answer.
+        if (after - first_sep != 2 or url[first_sep] != '/' or url[first_sep + 1] != '/')
+            return .{ .scheme = scheme, .authority = contestedOrAbsent(scheme, url[after..]) };
+        rest = url[after..];
+        delimited = true;
+    }
+
+    // A network-path reference names an authority and no scheme (RFC 3986
+    // §4.2). Without this the `/` cut below lands at byte zero, the authority is
+    // empty, and the line reports nothing at all — including for
+    // `//user:pa55@lists.example/x`, where the host is not in doubt.
+    // Exactly `//` here too, and for the same reason. RFC 3986 reads an authority
+    // after `//` and nothing else — `///hunter2/x` is an empty authority and the
+    // path `/hunter2/x`, and `\\hunter2\x` is a path outright. WHATWG resolves a
+    // reference against a base, and against a special-scheme base its
+    // ignore-slashes state reads `hunter2` as the host in both. A run that is not
+    // exactly `//` is therefore contested, not settled, and it is withheld rather
+    // than reported as an authority the url does not have.
+    //
+    // A run of one is settled: both readings make it a path. So is any run whose
+    // candidate authority is empty — `\\?\C:\lists\hosts.txt` ends the authority
+    // at its `?` under the reading that looks for one, so neither finds a host
+    // and there is nothing to contest.
+    if (!delimited) {
+        var run: usize = 0;
+        while (run < rest.len and isSeparator(rest[run])) run += 1;
+        if (run >= 2) {
+            if (candidateAuthority(rest[run..]).len == 0)
+                return .{ .scheme = scheme, .authority = "" };
+            if (run != 2 or rest[0] != '/' or rest[1] != '/')
+                return .{ .scheme = scheme, .authority = null };
+            rest = rest[2..];
+            delimited = true;
+        }
+    }
+
+    // The authority ends at the first `/`. Everything from there on is path,
+    // query or fragment, and none of the three is printed — so an `@` after
+    // that slash is not userinfo by any reading, and no longer needs to be one
+    // to stay out of the log.
+    const authority = rest[0 .. std.mem.indexOfScalar(u8, rest, '/') orelse rest.len];
+
+    // A scheme with no separator after it is an opaque path under RFC 3986 and,
+    // for a special scheme, a host under WHATWG — so `https:hunter2` is either a
+    // path segment, which is where a token lives, or a host. The disagreement is
+    // the whole of the evidence, exactly as it is for `https:a@hunter2`, and this
+    // check runs before the `@` lookup so both readings reach it.
+    //
+    // No exception is made for a suffix that looks like a port. An earlier
+    // revision took the digits in `localhost:8080` for one, which also let
+    // `https:123456` through, whose digits are an opaque path and as much a token
+    // as any other text. Neither is resolved now.
+    //
+    // Note what the two cases are not. `https:` is a WHATWG special scheme, so
+    // `https:hunter2` is contested — an opaque path to RFC 3986, a host to
+    // WHATWG. `localhost:` is not special, so `localhost:8080` is a scheme and an
+    // opaque path to *both*, and what an operator meant by it — a host and a
+    // port — is a reading no parser offers. It is withheld all the same, because
+    // the text after the colon is a path segment under every reading and a path
+    // segment is never printed.
+    //
+    // So the two get the same treatment and different answers: `https:hunter2` is
+    // withheld as contested, `localhost:8080` as naming no authority at all.
+    //
+    // An `IP:port` is unaffected: a leading digit fails the scheme production, so
+    // `10.0.0.2:8080` and `[::1]:853` resolve.
+    if (!delimited) {
+        const trimmed = cut(authority);
+        const colon = std.mem.indexOfScalar(u8, trimmed, ':') orelse trimmed.len;
+        if (colon != trimmed.len and isScheme(trimmed[0..colon])) return .{
+            .scheme = scheme,
+            .authority = contestedOrAbsent(trimmed[0..colon], trimmed[colon + 1 ..]),
+        };
+    }
+
+    // With no `@` there is no userinfo to remove and no side to choose. A `?`,
+    // a `#` or a `\` ends the authority; each of the three ends it under every
+    // reading of the text before it.
+    const at = std.mem.lastIndexOfScalar(u8, authority, '@') orelse
+        return .{ .scheme = scheme, .authority = cut(authority) };
+
+    // A `?`, a `#` or a `\` in front of that `@` makes the authority ambiguous,
+    // and the two readings put the host on opposite sides of the `@`. On
+    // `https://lists.example?token=prefix@hunter2` the text after it is an api
+    // key; on `https://user:pa55?@host` the text before it is a password. Both
+    // readings are available on both urls and nothing in either string tells
+    // them apart, so neither side may be printed.
+    if (std.mem.indexOfAny(u8, authority[0..at], "?#\\") != null)
+        return .{ .scheme = scheme, .authority = null };
+
+    // An `@` is a userinfo delimiter only inside an authority. The scan knows one
+    // is there in exactly three cases: a scheme delimiter introduced it, a `//`
+    // did, or the string names no scheme at all, where this file's rule is that
+    // the text before the first `/` is the closest thing to a host. The remaining
+    // case — a scheme with no separator after it — was withheld above, before the
+    // `@` was looked for, because it is ambiguous whether or not an `@` is in it.
+    return .{ .scheme = scheme, .authority = cut(authority[at + 1 ..]) };
+}
+
+
+/// `text` up to the first `?`, `#` or `\`, each of which ends an authority. A
+/// `\` is here rather than in `redact`'s `/` cut because the `/` cut runs before
+/// the userinfo is located and a `\` before an `@` is not a separator under
+/// every reading — `https://user:pa55\@host` is ambiguous, not hierarchical.
+fn cut(text: []const u8) []const u8 {
+    return text[0 .. std.mem.indexOfAny(u8, text, "?#\\") orelse text.len];
+}
+
+/// The `QuotedText` of `text`, which it borrows for the length of the call that
+/// prints it. The result prints its own quotes; see `QuotedText`.
+pub fn quoteText(text: []const u8) QuotedText {
+    return .{ .text = text };
+}
+
+/// The `redact` of `url`, quoted.
+///
+/// The rule for choosing between the two, so it does not have to be rediscovered
+/// per call site: **use this whenever anything follows the url on the line.** A
+/// redacted authority can still hold a space, a `:` and a `'`, so unquoted it can
+/// impersonate whatever comes next — `upstream {f} failed: {t}` with a url whose
+/// authority is `ok failed: Timeout` reports a failure that did not happen.
+///
+/// `redact` is for the two cases where that cannot arise: the url ends the line,
+/// or the caller owns the escaping for a delimiter of its own, as
+/// `web/metrics.zig` does for a Prometheus label value.
+///
+/// The result prints its own quotes; see `QuotedUrl`.
+pub fn redactQuoted(url: []const u8) QuotedUrl {
+    return .{ .url = redact(url) };
+}
+
+/// What ends a url's components. `/` is RFC 3986's. `\` is here because the
+/// strings this scan exists for are the ones a parser refuses:
+/// `https:\\dns.nextdns.io\abcd12` carries an account identifier after a `\`,
+/// and a scan that read only `/` printed the whole of it.
+///
+/// It ends a component; it does not open an authority. What that url *names* is
+/// contested — path text to RFC 3986, which gives `\` no meaning, and a host to
+/// WHATWG, which reads `\` as `/` for a special scheme — so `redact` withholds
+/// it. The `\` still matters here because both readings agree the account
+/// identifier after it is not part of any host.
+/// `null` when the readings disagree about whether `after` holds a host, empty
+/// when they agree it holds none. Both withhold; they differ in what the line
+/// claims, and `SafeUrl.authority` documents that as a real distinction.
+///
+/// Known imprecision, in the safe direction. Three shapes return `null` where
+/// both readings in fact find no host, so the line says "could not resolve"
+/// where "names none" is the truth:
+///
+///   - `https:/user@/x` — emptiness is decided before the userinfo is removed,
+///     so `user@` counts as a candidate host when the host after it is empty.
+///   - `\path@hunter2` — a leading run of one is settled as a path under both
+///     readings, but reaches the late-delimiter rule instead of returning here.
+///   - `file:secret` — `file` is in `special_schemes`, but WHATWG gives it its
+///     own parsing states in which that input is a local path with no host.
+///
+/// Each prints *less* than it could, never more; none prints the path segment.
+/// Fixing them means modelling more of two standards for inputs no accepted
+/// configuration can hold: every url this program takes carries `http`, `https`,
+/// `tls`, `udp` or `tcp` and a `//`, so all three shapes reach a line only as
+/// something the validator is already rejecting by field path.
+///
+/// Only a WHATWG special scheme reads an authority out of text a `//` did not
+/// introduce, so only a special scheme can disagree with RFC 3986 here. And
+/// nothing is contested when the reading that looks for a host finds none —
+/// `https:` alone names no authority under either.
+fn contestedOrAbsent(scheme: []const u8, after: []const u8) ?[]const u8 {
+    if (candidateAuthority(after).len == 0) return "";
+    return if (isSpecialScheme(scheme)) null else "";
+}
+
+/// The host a WHATWG-style reading would take out of `after`, used only to tell
+/// an empty one from a non-empty one.
+fn candidateAuthority(after: []const u8) []const u8 {
+    return cut(after[0 .. std.mem.indexOfScalar(u8, after, '/') orelse after.len]);
+}
+
+/// WHATWG's special schemes: the ones whose urls it reads an authority into
+/// without a `//`, and whose backslashes it converts to slashes.
+///
+/// This is a closed set fixed by the URL Standard, not a list of what this
+/// program supports. That is the difference between it and the known-scheme list
+/// an earlier revision removed: this one cannot go stale when nxdns learns a new
+/// transport, because it never described nxdns in the first place.
+const special_schemes = [_][]const u8{ "ftp", "file", "http", "https", "ws", "wss" };
+
+fn isSpecialScheme(scheme: []const u8) bool {
+    for (special_schemes) |special| {
+        if (std.ascii.eqlIgnoreCase(scheme, special)) return true;
+    }
+    return false;
+}
+
+const separators = "/\\";
+
+fn isSeparator(c: u8) bool {
+    return std.mem.indexOfScalar(u8, separators, c) != null;
+}
+
+/// Whether `text` is a url scheme: an ASCII letter followed by letters, digits,
+/// `+`, `-` and `.` — RFC 3986's production, which is what a scheme delimiter
+/// has to look like before the text in front of it may be dropped as one.
+fn isScheme(text: []const u8) bool {
+    if (text.len == 0) return false;
+    if (!std.ascii.isAlphabetic(text[0])) return false;
+    for (text[1..]) |c| {
+        if (std.ascii.isAlphanumeric(c)) continue;
+        if (c == '+' or c == '-' or c == '.') continue;
+        return false;
+    }
+    return true;
+}
+
+/// Writes `text` with every control character escaped, spending at most
+/// `budget` printed characters and never splitting an escape sequence. Returns
+/// whether the whole of `text` was written.
+///
+/// The control escapes are the ones `platform/logging.zig` uses on a whole log
+/// message, byte for byte, and that is deliberate. `delimiter` adds the one
+/// escape that sink has no reason to make.
+///
+/// Two layers escape the same bytes here and neither is redundant. Do not
+/// delete this one on the grounds that the log sink already covers it: that
+/// sink covers every `std.log` line and nothing else, and a redacted url does
+/// not only reach a sink. `config/validate.zig` builds `Problem.message` as an
+/// allocated string; `web/handlers/mutations.zig` prints that message into the
+/// body of the 400 from `POST /api/blocklists`, and `cli.zig` prints it to the
+/// stdout of an interactive `nxdns check`. Control bytes baked into that string
+/// reach an HTTP response and an operator's terminal, neither of which any log
+/// sink is in a position to escape. A url has to arrive safe rather than be
+/// made safe by whatever it is written to.
+///
+/// The cost is that a url inside a log line is escaped twice: `\n` in the row
+/// prints as `\\n` in journald and as `\n` on stdout. One notation across both
+/// is what keeps that legible.
+///
+/// A `\` is escaped for the same reason it is there: `\n` in the output then
+/// means the byte this function replaced and `\\n` means two characters an
+/// operator typed. It is what makes `\'` unambiguous as well.
+fn writeEscaped(
+    w: *std.Io.Writer,
+    text: []const u8,
+    budget: *usize,
+    delimiter: Delimiter,
+) std.Io.Writer.Error!bool {
+    const hex = "0123456789abcdef";
+    for (text) |byte| {
+        var hex_buf: [4]u8 = undefined;
+        const escape: ?[]const u8 = switch (byte) {
+            '\\' => "\\\\",
+            '\n' => "\\n",
+            '\r' => "\\r",
+            '\t' => "\\t",
+            '\'' => if (delimiter == .single_quote) "\\'" else null,
+            0x00...0x08, 0x0b, 0x0c, 0x0e...0x1f, 0x7f => blk: {
+                hex_buf = .{ '\\', 'x', hex[byte >> 4], hex[byte & 0x0f] };
+                break :blk &hex_buf;
+            },
+            else => null,
+        };
+        if (escape) |seq| {
+            if (budget.* < seq.len) return false;
+            try w.writeAll(seq);
+            budget.* -= seq.len;
+        } else {
+            if (budget.* < 1) return false;
+            try w.writeByte(byte);
+            budget.* -= 1;
+        }
+    }
+    return true;
+}
+
+/// The character the caller of `writeEscaped` wraps the escaped text in, which
+/// is therefore the one character beyond the control set that has to be escaped
+/// inside it. `.none` is a value nothing wraps: a `SafeUrl` is printed bare, and
+/// escaping a `'` in a host would say a `'` there means something it does not.
+const Delimiter = enum { none, single_quote };
+
+// ---------------------------------------------------------------------------
+// tests
+// ---------------------------------------------------------------------------
+
+const testing = std.testing;
+
+fn expectRedacted(expected: []const u8, url: []const u8) !void {
+    var buf: [8 * max_len]u8 = undefined;
+    try testing.expectEqualStrings(expected, try std.fmt.bufPrint(&buf, "{f}", .{redact(url)}));
+}
+
+test "redact drops the query, the fragment and the userinfo" {
+    // The credential shapes a source url can hold: an api key in the query, a
+    // signed url whose signature is a query parameter, and userinfo.
+    try expectRedacted(
+        "https://lists.example",
+        "https://lists.example/hosts.txt?apikey=s3cr3t",
+    );
+    try expectRedacted(
+        "https://cdn.example",
+        "https://cdn.example/l/hosts.txt?Expires=1700000000&Signature=abc123&Key-Pair-Id=K2",
+    );
+    try expectRedacted(
+        "https://lists.example",
+        "https://user:pa55@lists.example/hosts.txt",
+    );
+    try expectRedacted(
+        "https://lists.example:8443",
+        "https://token@lists.example:8443/hosts.txt?t=1#frag",
+    );
+}
+
+test "redact drops a credential carried in a path segment" {
+    // The path is a place a token lives — a per-subscriber download url is the
+    // common shape — and the rule against writing a secret to a log does not
+    // bend for the component it sits in.
+    try expectRedacted(
+        "https://lists.example",
+        "https://lists.example/download/token/hunter2/hosts.txt",
+    );
+    // NextDNS: the path segment is the account identifier, so this is the shape
+    // the rule exists for rather than an invented one.
+    try expectRedacted("https://dns.nextdns.io", "https://dns.nextdns.io/abcd12");
+    try expectRedacted("https://dns.nextdns.io", "https://dns.nextdns.io/abcd12/mydevice");
+    // Its DoT form puts the same id in the hostname, where redaction cannot
+    // reach it. Pinned so the limit stays visible rather than being discovered.
+    try expectRedacted("tls://abcd12.dns.nextdns.io", "tls://abcd12.dns.nextdns.io");
+    try expectRedacted("https://lists.example", "https://lists.example/hunter2");
+    // An `@` in the path is not userinfo, and no longer has to be told apart
+    // from it: the path goes either way.
+    try expectRedacted("https://lists.example", "https://lists.example/@who/hosts.txt");
+}
+
+test "redact keeps everything an operator needs to know where a source points" {
+    try expectRedacted("https://lists.example", "https://lists.example/hosts.txt");
+    try expectRedacted("http://10.0.0.2:8080", "http://10.0.0.2:8080/a/b.txt");
+    try expectRedacted("https://lists.example", "https://lists.example?apikey=s3cr3t");
+    // No host: what is left still names the scheme.
+    try expectRedacted("https://", "https://?apikey=s3cr3t");
+}
+
+test "redact holds on the malformed urls a parser refuses" {
+    // These reach the log through `error.BadUrl`, so scanning has to hold where
+    // `std.Uri.parse` gives up. A scheme delimiter of one separator, of three,
+    // and of none at all: each once carried the userinfo into the log, because
+    // the scan looked for `://` and found no authority without it.
+    //
+    // Each is now withheld rather than resolved. An earlier revision read the
+    // text after a run of any length as the authority, which drops the userinfo
+    // on these four but prints the path segment of `https:/hunter2` as a host.
+    // Only a run of exactly two says an authority is there; what these hold after
+    // one, or after three, is a path to RFC 3986 and a host to WHATWG. The
+    // property the line asserts is unchanged — no userinfo reaches the log — and
+    // it now holds by withholding rather than by resolving.
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/user:hunter2@host/list");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:///user:hunter2@host/list");
+    try expectRedacted("HTTPS://(ambiguous authority omitted)", "HTTPS:/user:hunter2@host/list");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:\\user:hunter2@host\\list");
+    // A separator of none at all is withheld for the same reason.
+    try expectRedacted("(ambiguous authority omitted)", "https:user:hunter2@host/list");
+    try expectRedacted("", "?apikey=s3cr3t");
+    try expectRedacted("not a url", "not a url");
+    try expectRedacted("", "");
+    // A colon that is not a scheme delimiter does not make one.
+    try expectRedacted("lists.example", "lists.example/a:/b");
+    try expectRedacted("", "/download/token/hunter2/hosts.txt");
+}
+
+test "redact treats a backslash as a hierarchical separator" {
+    // A url typed or pasted with backslashes reaches these lines through
+    // `error.BadUrl`, so the scan has to cut on one. NextDNS again, because the
+    // path segment it carries is the whole account identifier.
+    // A `\` ends an authority but never opens one. WHATWG converts it to a `/`
+    // only for a special scheme, and RFC 3986 gives it no meaning at all, so none
+    // of these names a host that both readings agree on — and `tls:\\host` has no
+    // reading at all that makes it one. The account identifier stays out of the
+    // line either way, which is the property this test is for.
+    try expectRedacted("https://(ambiguous authority omitted)", "https:\\\\dns.nextdns.io\\abcd12");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/\\dns.nextdns.io\\abcd12");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:\\/dns.nextdns.io\\abcd12");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:\\dns.nextdns.io\\abcd12");
+    try expectRedacted("tls://", "tls:\\\\abcd12.dns.nextdns.io");
+    // A token in a backslash path goes the way a token in a `/` path goes.
+    try expectRedacted(
+        "https://lists.example",
+        "https://lists.example\\download\\token\\hunter2\\hosts.txt",
+    );
+    // The `\\` does not open an authority, so this is reported by its scheme
+    // alone. What the assertion is really for is that the token in front of the
+    // `@` does not reach the line, and withholding delivers that at least as
+    // well as resolving did.
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https:\\\\token@lists.example:8443\\hosts.txt?t=1#frag",
+    );
+    // A backslash separator does not make a scheme out of a colon that is not
+    // one, exactly as a `/` does not.
+    try expectRedacted("lists.example", "lists.example\\a:\\b");
+    try expectRedacted("", "\\download\\token\\hunter2\\hosts.txt");
+}
+
+test "redact resolves a backslash authority only where one reading survives" {
+    // A `\` after the userinfo ends the authority under the WHATWG reading and
+    // is an illegal host byte under RFC 3986's, so both readings agree that
+    // nothing after it is a host. Cutting there prints less than either.
+    try expectRedacted("https://host", "https://user@host\\list");
+    // But only once a `//` has established that an authority is there at all. A
+    // leading `\\` does not, so the userinfo is withheld with everything else
+    // rather than cut out of a host that was never settled.
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https:\\\\user:hunter2@host\\list",
+    );
+
+    // A `\` in front of the `@` is the ambiguous shape instead, and used to
+    // print the text on one side of it: `https:\\lists.example\path@evil` gave
+    // `https://evil`. See "redact omits an authority it cannot resolve".
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https:\\\\lists.example\\path@evil",
+    );
+
+    // A Windows path is not a url and leaves nothing that names a host. The
+    // line it appears on still carries the source's row id and name.
+    try expectRedacted("", "\\\\?\\C:\\lists\\hosts.txt");
+}
+
+test "redact omits an authority it cannot resolve" {
+    // The shape this exists for: a query parameter whose value holds an `@`.
+    // The text after that `@` is the query, which is where an api key lives, and
+    // a scan that read it as the end of a userinfo printed the key where the
+    // host goes — having already dropped the real host.
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://lists.example?token=prefix@hunter2",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://lists.example?user=a@b.example&key=hunter2",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://lists.example#f@hunter2",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https:\\\\lists.example\\p@hunter2",
+    );
+
+    // The mirror image, which the previous ordering fixed and this keeps fixed:
+    // the text before the `@` is a credential just as often, so neither side may
+    // be printed. Both of these once printed `user:pa55` as the host.
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://user:pa55?@host/x",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://user:pa55#@host/x",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https://user:pa55\\@host/x",
+    );
+    try expectRedacted(
+        "https://(ambiguous authority omitted)",
+        "https:\\\\user:pa55?@host\\x",
+    );
+
+    // A url naming a scheme without a separator after it has no authority by
+    // RFC 3986 and one by the WHATWG parser, so the `@` in it is a path
+    // separator under the first reading and a userinfo delimiter under the
+    // second. `hunter2` is a path segment or a host and no scan can say which.
+    try expectRedacted("(ambiguous authority omitted)", "https:a@hunter2");
+    try expectRedacted("(ambiguous authority omitted)", "https:user:hunter2@host/list");
+    // `user` is not one of WHATWG's special schemes, so both readings make this a
+    // scheme and an opaque path and neither finds a host. It names no authority
+    // rather than one that cannot be resolved, and the password is withheld by
+    // the same rule either way.
+    try expectRedacted("", "user:pa55@lists.example/hosts.txt?apikey=s3cr3t");
+
+    // A url with no scheme at all prints the marker on its own, which is still
+    // not a url and still not a host.
+    try expectRedacted("(ambiguous authority omitted)", "lists.example?token=a@hunter2");
+}
+
+test "redact tells an omitted authority apart from an absent one" {
+    // Three outcomes an operator has to be able to tell apart, because the
+    // action each calls for differs: a host, no host, and a host this cannot
+    // name. Only the third withholds anything.
+    try expectRedacted("https://lists.example", "https://lists.example/hosts.txt");
+    try expectRedacted("https://", "https://?apikey=s3cr3t");
+    try expectRedacted("https://(ambiguous authority omitted)", "https://?u=a@hunter2");
+
+    try expectRedacted("lists.example", "lists.example/hosts.txt");
+    try expectRedacted("", "?apikey=s3cr3t");
+    try expectRedacted("(ambiguous authority omitted)", "?u=a@hunter2");
+}
+
+test "redact still resolves an authority whose delimiters follow the userinfo" {
+    // The ambiguity is an `@` after a `?`, a `#` or a `\`, not an `@` at all.
+    // Where the delimiters fall the way a url puts them, the host is not in
+    // doubt and withholding it would cost an operator the line's whole point.
+    try expectRedacted("https://lists.example", "https://user@lists.example?apikey=s3cr3t");
+    try expectRedacted("https://lists.example", "https://user:pa55@lists.example#frag");
+    try expectRedacted("https://lists.example:8443", "https://token@lists.example:8443\\hosts.txt");
+    try expectRedacted("https://lists.example", "https://user@lists.example/x?u=a@b");
+    // No `@` in the authority: a `?` still ends it, and nothing is ambiguous.
+    try expectRedacted("https://lists.example", "https://lists.example?apikey=s3cr3t");
+}
+
+test "redact escapes the control characters that would forge a log line" {
+    // A row can be written by a path that does not validate as strictly as the
+    // config validator, so the manager prints whatever the column holds. A
+    // newline in it would end the line and start one of the operator's
+    // choosing.
+    try expectRedacted(
+        "https://lists.example\\n2026-01-01 ERROR forged",
+        "https://lists.example\n2026-01-01 ERROR forged/hosts.txt",
+    );
+    try expectRedacted("https://a\\rb", "https://a\rb/x");
+    try expectRedacted("https://a\\tb", "https://a\tb/x");
+    try expectRedacted("https://a\\x00b", "https://a\x00b/x");
+    try expectRedacted("https://a\\x7fb", "https://a\x7fb/x");
+    // An ESC would reach a terminal as a control sequence on the one path the
+    // log sink does not cover: `cli.zig` prints a diagnostic to stdout.
+    try expectRedacted("https://a\\x1bb", "https://a\x1bb/x");
+    // A literal backslash ends the authority, so `redact` cannot print one at
+    // all. The doubling that keeps an escape sequence unambiguous is exercised
+    // where a backslash does survive: `quoteText`, below.
+    try expectRedacted("https://a", "https://a\\nb/x");
+}
+
+test "redact bounds the line it prints at max_len" {
+    const long_host = "h" ** (2 * max_len);
+    var buf: [8 * max_len]u8 = undefined;
+
+    const printed = try std.fmt.bufPrint(&buf, "{f}", .{redact("https://" ++ long_host ++ "/x")});
+    try testing.expectEqualStrings(("https://" ++ long_host)[0..max_len] ++ "...", printed);
+
+    // The bound counts what is printed, so an escape cannot spend four
+    // characters of a log line per byte of url.
+    const control_host = "\n" ** max_len;
+    const escaped = try std.fmt.bufPrint(&buf, "{f}", .{redact("https://" ++ control_host ++ "/x")});
+    try testing.expect(escaped.len <= max_len + 3);
+    try testing.expect(std.mem.endsWith(u8, escaped, "..."));
+    try testing.expect(!std.mem.containsAtLeast(u8, escaped, 1, "\n"));
+
+    // A scheme long enough on its own truncates inside the scheme rather than
+    // printing it whole and starting on the host.
+    const long_scheme = "s" ** (2 * max_len);
+    const truncated = try std.fmt.bufPrint(&buf, "{f}", .{redact(long_scheme ++ ":/host")});
+    try testing.expectEqualStrings(long_scheme[0..max_len] ++ "...", truncated);
+}
+
+test "quoteText escapes and bounds an operator-supplied name" {
+    var buf: [8 * max_len]u8 = undefined;
+
+    try testing.expectEqualStrings(
+        "'ads and trackers'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText("ads and trackers")}),
+    );
+    try testing.expectEqualStrings(
+        "'ads\\n2026-01-01 ERROR forged'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText("ads\n2026-01-01 ERROR forged")}),
+    );
+    // A name cannot close the quote around it and write what follows as though
+    // it were another field of the line.
+    try testing.expectEqualStrings(
+        "'ads\\' (https://decoy.example) --'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText("ads' (https://decoy.example) --")}),
+    );
+    // Nor by escaping the escape: a `\` before the quote is doubled first, so
+    // `\'` in the output is this function's and never the operator's.
+    try testing.expectEqualStrings(
+        "'ads\\\\\\' (https://decoy.example) --'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText("ads\\' (https://decoy.example) --")}),
+    );
+    // Nor by running past `max_len`: the truncation closes the quote too.
+    const long_name = "n" ** (2 * max_len);
+    try testing.expectEqualStrings(
+        "'" ++ long_name[0..max_len] ++ "...'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText(long_name)}),
+    );
+    try testing.expectEqualStrings(
+        "'" ++ "\\'" ** (max_len / 2) ++ "...'",
+        try std.fmt.bufPrint(&buf, "{f}", .{quoteText("'" ** max_len)}),
+    );
+}
+
+test "redact withholds a scheme whose separator is missing, with or without an @" {
+    // RFC 3986 reads `hunter2` as an opaque path and a path segment is where a
+    // token lives; WHATWG inserts the missing `//` for a special scheme and reads
+    // it as the host. An earlier revision withheld this only when an `@` was
+    // present, so the plainer shape printed the path whole.
+    try expectRedacted("(ambiguous authority omitted)", "https:hunter2");
+    try expectRedacted("(ambiguous authority omitted)", "https:a@hunter2");
+    try expectRedacted("(ambiguous authority omitted)", "https:123456");
+
+    // Only a WHATWG special scheme can disagree with RFC 3986 here, because only
+    // a special scheme reads an authority out of text no `//` introduced. For
+    // every other scheme both readings say the same thing — a scheme and an
+    // opaque path — so these name no authority rather than an unresolved one.
+    // Withheld either way; what differs is what the line then claims.
+    try expectRedacted("", "mailto:ops@example.com");
+    try expectRedacted("", "localhost:8080/hosts.txt");
+    try expectRedacted("", "localhost:8080@evil/x");
+
+    // An `IP:port` is unaffected, because a leading digit fails the scheme
+    // production and there is nothing to disagree about.
+    try expectRedacted("10.0.0.2:8080", "10.0.0.2:8080/hosts.txt");
+    try expectRedacted("[::1]:853", "[::1]:853/x");
+    try expectRedacted("lists.example", "lists.example/hosts.txt");
+}
+
+test "redact takes exactly two separators as an authority delimiter" {
+    // One separator leaves an absolute path: RFC 3986 reads `https:/hunter2` as
+    // the path `/hunter2`, WHATWG reads `hunter2` as the host. Three or more is
+    // an empty authority to the first and a host to the second. An earlier
+    // revision accepted any run and printed the path segment as the host.
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/hunter2");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:///hunter2");
+    try expectRedacted("localhost://", "localhost:/hunter2");
+    try expectRedacted("https://lists.example", "https://lists.example/hosts.txt");
+
+    // The userinfo this tolerance was introduced to protect is protected by
+    // withholding instead. The `/` cut runs ahead of the userinfo lookup now, so
+    // the authority of a url with no `://` ends before any `@` it holds.
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/user:pass@host/list");
+}
+
+test "redact resolves a network-path reference instead of reporting nothing" {
+    // The `/` cut lands at byte zero here, so without the `//` branch every one
+    // of these redacted to the empty string and the line named no source at all.
+    try expectRedacted("lists.example", "//lists.example/hosts.txt");
+    // Exactly two. A longer run is contested — an empty authority to RFC 3986,
+    // and `lists.example` as the host to WHATWG resolving against a
+    // special-scheme base — so it is withheld rather than reported as a url that
+    // names no authority. Those are different answers and this type keeps them
+    // apart.
+    try expectRedacted("(ambiguous authority omitted)", "///lists.example/x");
+    // The authority is not in doubt, so the userinfo is dropped rather than the
+    // whole of it withheld.
+    try expectRedacted("lists.example", "//user:pa55@lists.example/hosts.txt");
+    // An ambiguous one is still withheld: the branch says where the authority
+    // starts, not that every reading of it is settled.
+    try expectRedacted("(ambiguous authority omitted)", "//a?b@c");
+    // A single leading `/` is a path, not an authority, and still names none.
+    try expectRedacted("", "/path/only");
+}
+
+test "the shapes redact over-withholds on print less, never more" {
+    // The three imprecisions `contestedOrAbsent` documents. Each says "could not
+    // resolve" where both readings in fact find no host, so each prints less than
+    // it could. Pinned because the failure that matters is the other direction:
+    // if one of these ever starts naming a host, this test says so.
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/user@/x");
+    try expectRedacted("https://(ambiguous authority omitted)", "https:/user:hunter2@/x");
+    try expectRedacted("(ambiguous authority omitted)", "\\path@hunter2");
+    try expectRedacted("(ambiguous authority omitted)", "file:secret");
+    try expectRedacted("file://(ambiguous authority omitted)", "file:\\secret");
+    // `file` still resolves where a `//` settles it, which is why the imprecision
+    // is in the classification and not in the scan.
+    try expectRedacted("file://host", "file://host/secret");
+}
+
+test "redactQuoted writes its own quotes and closes them in every exit" {
+    var buf: [8 * max_len]u8 = undefined;
+
+    try testing.expectEqualStrings(
+        "'https://lists.example'",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://lists.example/hosts.txt")}),
+    );
+    // The omitted-authority value is prose with spaces in it, which is the case
+    // the quotes exist for: unquoted it runs into the words of the sentence
+    // around it.
+    try testing.expectEqualStrings(
+        "'https://(ambiguous authority omitted)'",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://lists.example?token=prefix@hunter2")}),
+    );
+    // A truncation returns early, and the closing quote still has to be written
+    // or the rest of the line reads as part of the value.
+    const long_host = "h" ** (2 * max_len);
+    try testing.expectEqualStrings(
+        "'https://" ++ ("h" ** (max_len - "https://".len)) ++ "...'",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://" ++ long_host ++ "/x")}),
+    );
+}
+
+test "a redacted authority cannot close the quote a caller would have added" {
+    var buf: [8 * max_len]u8 = undefined;
+
+    // A `'` is neither a component separator nor a control character, so it
+    // survives redaction into the authority. `redact` leaves it, which is why a
+    // caller may not supply the quotes itself.
+    try testing.expectEqualStrings(
+        "https://ho'st",
+        try std.fmt.bufPrint(&buf, "{f}", .{redact("https://ho'st/x")}),
+    );
+    try testing.expectEqualStrings(
+        "'https://ho\\'st'",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://ho'st/x")}),
+    );
+    // The forging shape, whole: an operator-supplied url that ends the value and
+    // writes what follows as though the line had said it.
+    try testing.expectEqualStrings(
+        "'https://ho\\' is fine; upstreams[9] '",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://ho' is fine; upstreams[9] /x")}),
+    );
+    // The escape-the-escape shape that `QuotedText` has to defend against does
+    // not arise here, and not because the escaper is different — it is the same
+    // one. A `\` ends an authority, so it never reaches the value to be doubled.
+    // This is asserted rather than assumed: it is the property that makes a
+    // single `\` in the output always this file's and never the operator's.
+    try testing.expectEqualStrings(
+        "'https://ho'",
+        try std.fmt.bufPrint(&buf, "{f}", .{redactQuoted("https://ho\\'st/x")}),
+    );
+}

src/server/udp_server.zig

@@ -77,7 +77,7 @@ pub const UdpServer = struct {
         /// handler, so nothing larger than 4096 bytes leaves this socket.
         ///
         /// Cost: 4096 + 65535 + the scratch below ≈ 74 KiB per slot, so the
-        /// default 64 slots hold ≈ 4.6 MiB. The PLAN §18 budget is 100 MB with
+        /// default 64 slots hold ≈ 4.6 MiB. The PLAN §18 budget is 100 MiB with
         /// ~1M blocked domains, so this pool takes about 5% of it.
         reply: [transport.max_message_len]u8,
         /// The handler's per-query working memory. It belongs to the slot so
@@ -297,7 +297,7 @@ test "a slot's reply buffer holds a whole DNS message" {
 
 test "the default slot pool stays inside the memory budget" {
     // 4096 + 65535 + scratch ≈ 74 KiB per slot; 64 slots ≈ 4.6 MiB, against the
-    // 100 MB of PLAN §18.
+    // 100 MiB of PLAN §18.
     const options: Options = .{};
     const pool_bytes = @sizeOf(UdpServer.Slot) * @as(usize, options.max_in_flight);
     try testing.expect(pool_bytes < 8 * 1024 * 1024);

src/storage/db.zig

@@ -1,12 +1,17 @@
 //! The whole SQLite surface nxdns owns (PLAN Decision G). Nothing above this
 //! file calls SQLite directly.
 //!
-//! **This file takes no `std.Io`.** It is the one deliberate exception to
-//! Decision E. SQLite performs its own file I/O through its VFS; routing it
-//! through `std.Io` would mean writing a custom SQLite VFS — a large,
+//! **SQLite's own file I/O takes no `std.Io`.** It is the one deliberate
+//! exception to Decision E. SQLite performs its own file I/O through its VFS;
+//! routing it through `std.Io` would mean writing a custom SQLite VFS — a large,
 //! security-sensitive component bought for nothing at household scale. Every
 //! other storage file that touches the filesystem takes `io: std.Io`.
 //!
+//! The exception covers SQLite, not nxdns. The one place this file does its own
+//! filesystem calls — the probes guarding `OpenMode.immutable`, at the open and
+//! again at `Db.verifyImmutable` — follows the ordinary rule and takes an
+//! `io: std.Io`, which is why that mode carries one.
+//!
 //! The C API is declared by hand below. No `@cImport` — the handles stay
 //! opaque, matching `src/platform/tls_server.zig`'s Mbed TLS approach.
 
@@ -135,6 +140,11 @@ pub const Error = error{
     SqliteError,
     OutOfMemory,
     Unexpected,
+    /// Not a SQLite result code. An `OpenMode.immutable` read would have
+    /// answered from a stale main file, because `<path>-wal` holds bytes or the
+    /// files moved while the read ran. Raised by the open and again by
+    /// `Db.verifyImmutable`. See `OpenMode.immutable`.
+    WalPending,
 };
 
 /// Maps a primary SQLite result code to `Error`. `SQLITE_NOMEM` becomes
@@ -190,19 +200,73 @@ fn check(code: c_int) Error!void {
     return mapCode(code);
 }
 
-pub const OpenMode = enum { read_write_create, read_write_existing, read_only, memory };
+pub const OpenMode = union(enum) {
+    read_write_create,
+    read_write_existing,
+    read_only,
+    memory,
+    /// Read a database that this process promises not to change, and that no
+    /// writer may be touching: `SQLITE_OPEN_READONLY` plus the `immutable=1` URI
+    /// parameter, which makes the pager treat the file like a temp file — no
+    /// locking, no rollback journal, no wal-index — so **no `-wal` and no `-shm`
+    /// appear beside it**.
+    ///
+    /// This mode exists for `nxdns check` (milestone-13 ruling F-c), which must
+    /// validate without writing. `.read_only` alone is not enough, and this is
+    /// measured, not assumed: reading a database whose header says WAL makes
+    /// SQLite build the wal-index, so `config.db-wal` (0 bytes) and
+    /// `config.db-shm` (32 KiB) are created, and a read-only connection cannot
+    /// remove them on close. A command that claims to write nothing must not
+    /// leave two files behind. Do not "simplify" this back to `.read_only`.
+    ///
+    /// What `immutable=1` costs: SQLite then **ignores any `-wal` file**. The
+    /// newest committed rows live there, so an immutable read of a database with
+    /// an un-checkpointed WAL would answer from stale data and say nothing — a
+    /// worse failure than the two sidecar files it removes. `Db.open` therefore
+    /// refuses this mode with `error.WalPending` whenever `<path>-wal` exists and
+    /// is not empty; the caller reports that as a failure and names `nxdns run`,
+    /// which opens read-write and checkpoints, as the fix.
+    ///
+    /// The guard lives inside `open` rather than in a helper callers are trusted
+    /// to call, because the failure it prevents is silent: a caller that forgets
+    /// a helper gets a plausible wrong answer, and nothing anywhere reports it.
+    ///
+    /// **The open is half the guard.** `immutable=1` takes no lock, so nothing
+    /// keeps a writer out for the duration of the read, and a check made only
+    /// before the read can say only that the log was empty *then*.
+    /// `Db.verifyImmutable` makes the other half, and a caller that grades what
+    /// it read without calling it is back to the stale answer this mode exists
+    /// to refuse.
+    ///
+    /// A zero-length `-wal` does not block the open: it holds no frames, so the
+    /// main file is complete. That is the exact leftover the pre-F-c `check`
+    /// used to create.
+    ///
+    /// The `std.Io` is for that probe — the one filesystem call nxdns itself
+    /// makes in this file. Passing it is what makes the guard unskippable.
+    immutable: std.Io,
+};
 
 pub const OpenOptions = struct {
     mode: OpenMode = .read_write_create,
     busy_timeout_ms: c_int = 5000,
 };
 
+/// SQLite's name for the write-ahead log beside `<path>`. Exported so a caller
+/// reporting `error.WalPending` can name the file without hard-coding SQLite's
+/// naming convention, which this file owns.
+pub const wal_suffix = "-wal";
+
 /// One SQLite connection.
 ///
 /// A `Db` must not move once a `Stmt` prepared from it is alive: every `Stmt`
 /// holds a `*Db`.
 pub const Db = struct {
     handle: *c.Sqlite3,
+    /// Set by `OpenMode.immutable` and null in every other mode: what the
+    /// database file and its `-wal` looked like when the read began, for
+    /// `verifyImmutable` to compare against when it ends.
+    immutable_guard: ?ImmutableGuard = null,
 
     /// Every mode carries `FULLMUTEX` (serialized mode). Phase 6's query logger
     /// and Phase 8's API handlers share one handle across `std.Io` tasks, and a
@@ -220,47 +284,59 @@ pub const Db = struct {
             .read_write_create, .memory => base | open_flag.readwrite | open_flag.create,
             .read_write_existing => base | open_flag.readwrite,
             .read_only => base | open_flag.readonly,
+            // Its own function: the URI buffer is 12 KiB, and every other open
+            // in the process would carry it in this frame.
+            .immutable => |io| return openImmutable(io, path, options.busy_timeout_ms),
         };
         const filename: [:0]const u8 = switch (options.mode) {
             .memory => ":memory:",
             else => path,
         };
 
-        var handle: ?*c.Sqlite3 = null;
-        const rc = c.sqlite3_open_v2(filename.ptr, &handle, flags, null);
-        if (rc != result.ok) {
-            // sqlite3_open_v2 allocates a handle even on failure. Read the
-            // message from it, then close it; dropping it leaks on every
-            // failed open.
-            // Logged at `warn`, not `err`: the failure itself reaches the
-            // caller as a typed error, and this line only carries the message
-            // that would otherwise die with the handle.
-            if (handle) |h| {
-                log.warn("sqlite3_open_v2 failed for '{s}': {s} (code {d}/{d})", .{
-                    filename,
-                    std.mem.span(c.sqlite3_errmsg(h)),
-                    rc & 0xff,
-                    c.sqlite3_extended_errcode(h),
-                });
-                _ = c.sqlite3_close_v2(h);
-            } else {
-                log.warn("sqlite3_open_v2 failed for '{s}': {s} (code {d})", .{
-                    filename,
-                    std.mem.span(c.sqlite3_errstr(rc)),
-                    rc,
-                });
-            }
-            return mapCode(rc);
-        }
-        const h = handle orelse return error.SqliteError;
+        const h = try openHandle(filename, flags);
+        return applyBusyTimeout(h, options.busy_timeout_ms);
+    }
 
-        // A silently ignored busy timeout is how a contended WAL database turns
-        // into random SQLITE_BUSY failures under load.
-        check(c.sqlite3_busy_timeout(h, options.busy_timeout_ms)) catch |e| {
-            _ = c.sqlite3_close_v2(h);
-            return e;
-        };
-        return .{ .handle = h };
+    /// Proves that the files an `OpenMode.immutable` read answered from stood
+    /// still while it ran, and fails the read with `error.WalPending` when they
+    /// did not. Call it after the last read and before anything read is
+    /// reported.
+    ///
+    /// `immutable=1` takes no lock at all — that is what stops the pager
+    /// building a wal-index, and the price is that nothing keeps a writer out.
+    /// The probe at open time can only say the log was empty at that instant: a
+    /// writer that appends one frame the instant after leaves the read answering
+    /// from the older pages of the main file, silently, which is the whole
+    /// failure `OpenMode.immutable`'s guard exists to prevent.
+    ///
+    /// Two things are compared, because a writer can hide in either:
+    ///
+    /// - `<path>-wal` holding bytes now. A log that appeared, one that grew, and
+    ///   one written into the empty file the open accepted all land here.
+    /// - `<path>` itself moving — size, inode, mtime or ctime. This is the
+    ///   checkpoint the log cannot show: a writer that checkpointed into the main
+    ///   file and truncated its log back to nothing leaves both stats saying "no
+    ///   frames" while the pages the read saw have been replaced.
+    ///
+    /// Best effort, and honestly so: a filesystem with coarse timestamps can
+    /// hide a rewrite that lands on the same byte count in the same tick. That
+    /// cannot be fixed from outside SQLite's locking, and taking a lock is the
+    /// one thing this mode may not do. What it closes is the window a single
+    /// stat before the read leaves open for the whole of the read.
+    ///
+    /// Two stats and nothing else: no `-wal` or `-shm` is created, and neither
+    /// sidecar is removed or truncated. Calling this on a handle opened in any
+    /// other mode is a caller bug.
+    pub fn verifyImmutable(self: *Db) Error!void {
+        const guard = self.immutable_guard orelse unreachable;
+        const now = try markImmutable(guard.io, guard.path);
+        if (!std.meta.eql(now, guard.mark)) {
+            log.warn(
+                "'{s}' changed while it was being read without a lock; the read is not trustworthy",
+                .{guard.path},
+            );
+            return error.WalPending;
+        }
     }
 
     pub fn close(self: *Db) void {
@@ -355,6 +431,199 @@ pub const Db = struct {
     }
 };
 
+fn openHandle(filename: [:0]const u8, flags: c_int) Error!*c.Sqlite3 {
+    var handle: ?*c.Sqlite3 = null;
+    const rc = c.sqlite3_open_v2(filename.ptr, &handle, flags, null);
+    if (rc != result.ok) {
+        // sqlite3_open_v2 allocates a handle even on failure. Read the
+        // message from it, then close it; dropping it leaks on every
+        // failed open.
+        // Logged at `warn`, not `err`: the failure itself reaches the
+        // caller as a typed error, and this line only carries the message
+        // that would otherwise die with the handle.
+        if (handle) |h| {
+            log.warn("sqlite3_open_v2 failed for '{s}': {s} (code {d}/{d})", .{
+                filename,
+                std.mem.span(c.sqlite3_errmsg(h)),
+                rc & 0xff,
+                c.sqlite3_extended_errcode(h),
+            });
+            _ = c.sqlite3_close_v2(h);
+        } else {
+            log.warn("sqlite3_open_v2 failed for '{s}': {s} (code {d})", .{
+                filename,
+                std.mem.span(c.sqlite3_errstr(rc)),
+                rc,
+            });
+        }
+        return mapCode(rc);
+    }
+    return handle orelse error.SqliteError;
+}
+
+/// A silently ignored busy timeout is how a contended WAL database turns into
+/// random SQLITE_BUSY failures under load.
+fn applyBusyTimeout(h: *c.Sqlite3, busy_timeout_ms: c_int) Error!Db {
+    check(c.sqlite3_busy_timeout(h, busy_timeout_ms)) catch |e| {
+        _ = c.sqlite3_close_v2(h);
+        return e;
+    };
+    return .{ .handle = h };
+}
+
+// ---------------------------------------------------------------------------
+// OpenMode.immutable
+// ---------------------------------------------------------------------------
+
+const uri_scheme = "file:";
+const uri_immutable_query = "?immutable=1";
+
+/// Worst case: every byte of the longest path the platform accepts becomes
+/// `%HH`.
+const immutable_uri_buf_len =
+    uri_scheme.len + 3 * std.Io.Dir.max_path_bytes + uri_immutable_query.len + 1;
+
+fn openImmutable(io: std.Io, path: [:0]const u8, busy_timeout_ms: c_int) Error!Db {
+    const mark = try markImmutable(io, path);
+
+    var buf: [immutable_uri_buf_len]u8 = undefined;
+    const uri = try immutableUri(&buf, path);
+
+    // `uri` without `open_flag.uri` would be opened as a filename spelled
+    // "file:...", creating nothing and finding nothing.
+    const flags = open_flag.exrescode | open_flag.fullmutex |
+        open_flag.readonly | open_flag.uri;
+    const h = try openHandle(uri, flags);
+    var database = try applyBusyTimeout(h, busy_timeout_ms);
+    database.immutable_guard = .{ .io = io, .path = path, .mark = mark };
+    return database;
+}
+
+/// What `OpenMode.immutable` recorded at the start of a read so that
+/// `Db.verifyImmutable` can prove nothing moved by the end of it.
+pub const ImmutableGuard = struct {
+    io: std.Io,
+    /// Borrowed. Must outlive the `Db`, which every caller satisfies by owning
+    /// the path for at least as long as the connection it opened with it.
+    path: []const u8,
+    mark: FileMark,
+};
+
+/// The main database file at one instant, in the fields an outside observer can
+/// compare cheaply. `atime` is deliberately absent: reading the file changes it,
+/// so comparing it would report every read as a change.
+///
+/// All zero when the file does not exist, which is itself a state worth
+/// comparing — a database replaced by an unlink is a database that moved.
+const FileMark = struct {
+    present: bool,
+    size: u64,
+    inode: std.Io.File.INode,
+    mtime_ns: i96,
+    ctime_ns: i96,
+};
+
+/// The state an immutable read must find unchanged, or `error.WalPending` when
+/// `<path>-wal` already holds bytes.
+///
+/// The `-wal` rule is deliberately conservative: any non-empty log fails.
+/// Deciding whether it really holds committed frames means running WAL recovery
+/// — checksums, salt, the wal-index — which is the writing that
+/// `OpenMode.immutable` exists to avoid. A live writer, a crash, and a
+/// checkpointed-but-retained log all land here, and refusing to answer is the
+/// right side to err on: the alternative is a stale answer nobody can see is
+/// stale. A zero-length log holds no frames, so the main file is complete and it
+/// passes.
+///
+/// A failed stat is not "no WAL": it means this cannot be known, so it stays a
+/// failure.
+fn markImmutable(io: std.Io, path: []const u8) Error!FileMark {
+    var buf: [std.Io.Dir.max_path_bytes + wal_suffix.len]u8 = undefined;
+    const sidecar = std.fmt.bufPrint(&buf, "{s}{s}", .{ path, wal_suffix }) catch
+        return error.TooBig;
+
+    if (try statOrAbsent(io, sidecar)) |wal| {
+        if (wal.size > 0) return error.WalPending;
+    }
+
+    const main = try statOrAbsent(io, path) orelse return .{
+        .present = false,
+        .size = 0,
+        .inode = 0,
+        .mtime_ns = 0,
+        .ctime_ns = 0,
+    };
+    return .{
+        .present = true,
+        .size = main.size,
+        .inode = main.inode,
+        .mtime_ns = main.mtime.nanoseconds,
+        .ctime_ns = main.ctime.nanoseconds,
+    };
+}
+
+fn statOrAbsent(io: std.Io, path: []const u8) Error!?std.Io.Dir.Stat {
+    return std.Io.Dir.cwd().statFile(io, path, .{}) catch |e| switch (e) {
+        error.FileNotFound => return null,
+        else => {
+            log.warn("cannot stat '{s}': {t}", .{ path, e });
+            return error.IoErr;
+        },
+    };
+}
+
+/// `file:` + percent-encoded `path` + `?immutable=1`.
+///
+/// The encoding is load-bearing, not cosmetic. `?` opens SQLite's query section
+/// and `#` its fragment, so an unencoded data directory named `dns?db` would
+/// silently open a *different* file; `%` must be encoded because SQLite decodes
+/// `%HH` on its way back to a filename. `--data-dir` is operator input, so all
+/// three are reachable.
+///
+/// Everything outside the unreserved set (`A-Z a-z 0-9 - . _ ~ /`) is encoded,
+/// which is always safe: SQLite decodes every escape in the path before handing
+/// the name to its VFS, so the bytes it opens are the bytes passed in.
+///
+/// `/` stays literal to keep diagnostics readable, with one exception. SQLite
+/// reads `file://…` as a URI authority and rejects any authority but the empty
+/// one or `localhost` (`sqlite3ParseUri`), so a path beginning with `//` — legal
+/// POSIX — has its second slash encoded.
+fn immutableUri(buf: []u8, path: []const u8) error{TooBig}![:0]const u8 {
+    var out: usize = 0;
+    try appendSlice(buf, &out, uri_scheme);
+    for (path, 0..) |ch, i| {
+        const opens_authority = i == 1 and ch == '/' and path[0] == '/';
+        if (isUriUnreserved(ch) and !opens_authority) {
+            try appendByte(buf, &out, ch);
+        } else {
+            const hex = "0123456789ABCDEF";
+            try appendByte(buf, &out, '%');
+            try appendByte(buf, &out, hex[ch >> 4]);
+            try appendByte(buf, &out, hex[ch & 0xf]);
+        }
+    }
+    try appendSlice(buf, &out, uri_immutable_query);
+    try appendByte(buf, &out, 0);
+    return buf[0 .. out - 1 :0];
+}
+
+fn isUriUnreserved(ch: u8) bool {
+    return switch (ch) {
+        'a'...'z', 'A'...'Z', '0'...'9', '-', '.', '_', '~', '/' => true,
+        else => false,
+    };
+}
+
+fn appendByte(buf: []u8, out: *usize, ch: u8) error{TooBig}!void {
+    if (out.* == buf.len) return error.TooBig;
+    buf[out.*] = ch;
+    out.* += 1;
+}
+
+fn appendSlice(buf: []u8, out: *usize, bytes: []const u8) error{TooBig}!void {
+    for (bytes) |ch| try appendByte(buf, out, ch);
+}
+
 /// One prepared statement.
 ///
 /// There is deliberately **no prepared-statement cache in this milestone**.
@@ -735,6 +1004,235 @@ test "a row-producing statement reports its row through step" {
     try testing.expect(try stmt.step());
 }
 
+// ---------------------------------------------------------------------------
+// OpenMode.immutable
+// ---------------------------------------------------------------------------
+
+/// `std.testing.tmpDir` creates its directory under `.zig-cache/tmp/` relative to
+/// the process working directory, which is also how SQLite's VFS resolves the
+/// filename it is handed (`queries_repo.zig:721`).
+const tmp_prefix = ".zig-cache/tmp/";
+const sub_path_len = @typeInfo(@FieldType(testing.TmpDir, "sub_path")).array.len;
+const test_io = testing.io;
+
+fn tmpPath(buf: []u8, tmp: *const testing.TmpDir, name: []const u8) ![:0]const u8 {
+    return std.fmt.bufPrintZ(buf, "{s}{s}/{s}", .{ tmp_prefix, &tmp.sub_path, name });
+}
+
+/// A file database in WAL mode holding one row, `id = marker`. Closing the last
+/// connection checkpoints and unlinks both sidecars, but the header keeps saying
+/// WAL — which is what makes a later `.read_only` open recreate them.
+fn writeWalDatabase(path: [:0]const u8, marker: i64) !void {
+    var database = try Db.open(path, .{ .mode = .read_write_create });
+    defer database.close();
+    try applyPragmas(&database, .{});
+    try database.exec("CREATE TABLE t (id INTEGER PRIMARY KEY);");
+    var stmt = try database.prepare("INSERT INTO t (id) VALUES (?1)");
+    defer stmt.deinit();
+    try stmt.bindInt(1, marker);
+    try stmt.exec();
+}
+
+/// A second writer doing what a running nxdns does: it appends to the
+/// write-ahead log and, being the last connection, checkpoints into the main
+/// file and unlinks both sidecars on close. The blob is what makes the main
+/// file grow by whole pages, so a test that watches for the change does not rest
+/// on the filesystem's timestamp resolution.
+fn checkpointOver(path: [:0]const u8) !void {
+    var database = try Db.open(path, .{ .mode = .read_write_existing });
+    defer database.close();
+    try applyPragmas(&database, .{});
+    try database.exec("INSERT INTO t (id) VALUES (8);");
+    try database.exec("CREATE TABLE bulk (v TEXT);");
+    try database.exec("INSERT INTO bulk (v) VALUES (hex(randomblob(30000)));");
+}
+
+fn expectAbsent(dir: std.Io.Dir, name: []const u8) !void {
+    dir.access(test_io, name, .{}) catch |e| switch (e) {
+        error.FileNotFound => return,
+        else => |other| return other,
+    };
+    std.debug.print("sidecar '{s}' exists and must not\n", .{name});
+    return error.SidecarPresent;
+}
+
+test "immutableUri encodes what would otherwise change which file is opened" {
+    var buf: [256]u8 = undefined;
+
+    try testing.expectEqualStrings(
+        "file:/var/lib/nxdns/config.db?immutable=1",
+        try immutableUri(&buf, "/var/lib/nxdns/config.db"),
+    );
+    // '?' would start SQLite's query section, '#' its fragment, '%' an escape.
+    try testing.expectEqualStrings(
+        "file:/data%3Fdir/config.db?immutable=1",
+        try immutableUri(&buf, "/data?dir/config.db"),
+    );
+    try testing.expectEqualStrings(
+        "file:/data%23dir/config.db?immutable=1",
+        try immutableUri(&buf, "/data#dir/config.db"),
+    );
+    try testing.expectEqualStrings(
+        "file:/data%25dir/config.db?immutable=1",
+        try immutableUri(&buf, "/data%dir/config.db"),
+    );
+    try testing.expectEqualStrings(
+        "file:/a%20b/c%3Fd%23e%25f.db?immutable=1",
+        try immutableUri(&buf, "/a b/c?d#e%f.db"),
+    );
+    // A relative path stays relative: SQLite's VFS resolves it against the
+    // working directory, exactly as a bare filename would be.
+    try testing.expectEqualStrings(
+        "file:config.db?immutable=1",
+        try immutableUri(&buf, "config.db"),
+    );
+    // A leading "//" would be read as a URI authority and rejected.
+    try testing.expectEqualStrings(
+        "file:/%2Fnet/share/config.db?immutable=1",
+        try immutableUri(&buf, "//net/share/config.db"),
+    );
+    // Only the authority position is special: "//" further in stays literal.
+    try testing.expectEqualStrings(
+        "file:/net//share/config.db?immutable=1",
+        try immutableUri(&buf, "/net//share/config.db"),
+    );
+    // Non-ASCII bytes survive the round trip because SQLite decodes them back.
+    try testing.expectEqualStrings(
+        "file:/caf%C3%A9/config.db?immutable=1",
+        try immutableUri(&buf, "/café/config.db"),
+    );
+
+    var small: [16]u8 = undefined;
+    try testing.expectError(error.TooBig, immutableUri(&small, "/var/lib/nxdns/config.db"));
+}
+
+test "an immutable open of a WAL database creates no -wal and no -shm" {
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    const path = try tmpPath(&path_buf, &tmp, "config.db");
+
+    try writeWalDatabase(path, 7);
+    try expectAbsent(tmp.dir, "config.db-wal");
+    try expectAbsent(tmp.dir, "config.db-shm");
+
+    {
+        var database = try Db.open(path, .{ .mode = .{ .immutable = test_io } });
+        defer database.close();
+        try testing.expectEqual(@as(i64, 7), try database.queryInt("SELECT id FROM t"));
+        // The point of the mode: a `.read_only` open creates both of these here,
+        // and cannot delete them on close.
+        try expectAbsent(tmp.dir, "config.db-wal");
+        try expectAbsent(tmp.dir, "config.db-shm");
+    }
+    try expectAbsent(tmp.dir, "config.db-wal");
+    try expectAbsent(tmp.dir, "config.db-shm");
+}
+
+test "an immutable open of a path holding URI metacharacters opens the intended file" {
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var decoy_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    var path_buf: [tmp_prefix.len + sub_path_len + 64]u8 = undefined;
+
+    // Unencoded, SQLite cuts the filename at the '?' and opens "<tmp>/d". That
+    // file exists here and holds a different database, so the failure without
+    // percent-encoding is a wrong answer, not an error.
+    try tmp.dir.createDirPath(test_io, "d?x#y%z");
+    try writeWalDatabase(try tmpPath(&decoy_buf, &tmp, "d"), 99);
+    const path = try tmpPath(&path_buf, &tmp, "d?x#y%z/config.db");
+    try writeWalDatabase(path, 7);
+
+    var database = try Db.open(path, .{ .mode = .{ .immutable = test_io } });
+    defer database.close();
+    try testing.expectEqual(@as(i64, 7), try database.queryInt("SELECT id FROM t"));
+}
+
+test "an immutable open refuses a database whose -wal holds bytes" {
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    const path = try tmpPath(&path_buf, &tmp, "config.db");
+    try writeWalDatabase(path, 7);
+
+    // What an unclean shutdown leaves behind, written directly so the case does
+    // not depend on when SQLite decides to checkpoint.
+    try tmp.dir.writeFile(test_io, .{ .sub_path = "config.db" ++ wal_suffix, .data = &([_]u8{0x37} ** 32) });
+    try testing.expectError(error.WalPending, Db.open(path, .{ .mode = .{ .immutable = test_io } }));
+
+    // A zero-length `-wal` holds no frames, so the main file is complete: the
+    // exact leftover the pre-F-c `check` created must not block a check.
+    try tmp.dir.writeFile(test_io, .{ .sub_path = "config.db" ++ wal_suffix, .data = "" });
+    var database = try Db.open(path, .{ .mode = .{ .immutable = test_io } });
+    defer database.close();
+    try testing.expectEqual(@as(i64, 7), try database.queryInt("SELECT id FROM t"));
+}
+
+test "an immutable read refuses a -wal that arrives while it is in flight" {
+    // The probe at open time can only say the log was empty *then*.
+    // `immutable=1` takes no lock, so a writer is free to arrive one instant
+    // later, and the read goes on answering from the older pages of the main
+    // file with nothing anywhere reporting it.
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    const path = try tmpPath(&path_buf, &tmp, "config.db");
+    try writeWalDatabase(path, 7);
+
+    var database = try Db.open(path, .{ .mode = .{ .immutable = test_io } });
+    defer database.close();
+    try testing.expectEqual(@as(i64, 7), try database.queryInt("SELECT id FROM t"));
+    // Nothing has moved yet, so the read stands.
+    try database.verifyImmutable();
+
+    // A zero-length log still holds no frames: the rule at the end of the read
+    // is the rule at the start of it.
+    try tmp.dir.writeFile(test_io, .{ .sub_path = "config.db" ++ wal_suffix, .data = "" });
+    try database.verifyImmutable();
+
+    // Frames, now, in the log the open accepted as empty.
+    try tmp.dir.writeFile(test_io, .{
+        .sub_path = "config.db" ++ wal_suffix,
+        .data = &([_]u8{0x37} ** 32),
+    });
+    try testing.expectError(error.WalPending, database.verifyImmutable());
+    // Repeatable: reporting the race is all it does.
+    try testing.expectError(error.WalPending, database.verifyImmutable());
+
+    // And it repairs nothing. The operator's log is byte for byte what was
+    // written, and no wal-index appeared beside it.
+    const wal = try tmp.dir.statFile(test_io, "config.db" ++ wal_suffix, .{});
+    try testing.expectEqual(@as(u64, 32), wal.size);
+    try expectAbsent(tmp.dir, "config.db-shm");
+}
+
+test "an immutable read refuses a main file checkpointed under it" {
+    // The case a `-wal` probe cannot see at either end: a writer checkpointed
+    // into the main file and, closing, unlinked its log again. Both stats say
+    // "no frames" while the pages the read answered from have been replaced.
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    const path = try tmpPath(&path_buf, &tmp, "config.db");
+    try writeWalDatabase(path, 7);
+
+    var database = try Db.open(path, .{ .mode = .{ .immutable = test_io } });
+    defer database.close();
+    try testing.expectEqual(@as(i64, 7), try database.queryInt("SELECT id FROM t"));
+    try database.verifyImmutable();
+
+    const before = try tmp.dir.statFile(test_io, "config.db", .{});
+    try checkpointOver(path);
+    const after = try tmp.dir.statFile(test_io, "config.db", .{});
+
+    // The premise of the test, proved rather than assumed: the main file really
+    // did move, and the log really is gone again.
+    try testing.expect(after.size != before.size);
+    try expectAbsent(tmp.dir, "config.db" ++ wal_suffix);
+
+    try testing.expectError(error.WalPending, database.verifyImmutable());
+}
+
 test "a duplicate insert into a UNIQUE column returns error.Constraint" {
     var db = try openMemory();
     defer db.close();

src/storage/migrations.zig

@@ -31,6 +31,9 @@ const ddl_v2: [:0]const u8 =
     \\ALTER TABLE upstreams ADD COLUMN tls_name TEXT NOT NULL DEFAULT '';
 ;
 
+/// The schema version this binary expects. A database `readVersion` reports
+/// below this needs `nxdns run` to migrate it; above it is `error.SchemaTooNew`
+/// and needs a newer nxdns.
 pub const target_version: u32 = steps[steps.len - 1].version;
 
 comptime {
@@ -102,9 +105,16 @@ pub fn migrateSteps(database: *db.Db, list: []const Step) Error!u32 {
     return target;
 }
 
+/// The schema version stamped in `database`, compared against `target_version`.
+///
 /// `0` when `schema_version` does not exist yet. Zero rows or more than one row
 /// is `error.SchemaCorrupt` — the version of a database is never guessed.
-fn readVersion(database: *db.Db) Error!u32 {
+///
+/// Reads only, so it works on a connection opened `.read_only` or
+/// `.immutable`. That is what it is public for: `nxdns check` may not migrate
+/// (ruling F-c), and "at version 1, this binary expects 2" tells an operator
+/// what to do where a bare SQLite error message does not.
+pub fn readVersion(database: *db.Db) Error!u32 {
     const present = try database.queryInt(
         "SELECT count(*) FROM sqlite_schema WHERE type='table' AND name='schema_version'",
     );
@@ -302,6 +312,47 @@ test "a failing step after step 2 rolls back the whole upgrade from version 1" {
     try testing.expectEqual(@as(u32, 1), try readVersion(&database));
 }
 
+test "readVersion reports 0 before a migration and target_version after it" {
+    var database = try openMigrated();
+    defer database.close();
+
+    try testing.expectEqual(@as(u32, 0), try readVersion(&database));
+    _ = try migrate(&database);
+    try testing.expectEqual(target_version, try readVersion(&database));
+}
+
+/// `.zig-cache/tmp/` is where `std.testing.tmpDir` puts its directories, and
+/// SQLite's VFS resolves filenames against the same working directory
+/// (`db.zig`'s immutable tests).
+const tmp_prefix = ".zig-cache/tmp/";
+const sub_path_len = @typeInfo(@FieldType(testing.TmpDir, "sub_path")).array.len;
+
+test "readVersion reads a file database through an immutable open, writing nothing" {
+    var tmp = testing.tmpDir(.{});
+    defer tmp.cleanup();
+    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
+    const path = try std.fmt.bufPrintZ(&path_buf, "{s}{s}/config.db", .{ tmp_prefix, &tmp.sub_path });
+
+    // A database an older nxdns left at version 1. `check` must report that, not
+    // migrate it (ruling F-c).
+    {
+        var database = try db.Db.open(path, .{ .mode = .read_write_create });
+        defer database.close();
+        try db.applyPragmas(&database, .{});
+        const first = [_]Step{.{ .version = 1, .sql = config_schema.ddl_v1 }};
+        try testing.expectEqual(@as(u32, 1), try migrateSteps(&database, &first));
+    }
+
+    var database = try db.Db.open(path, .{ .mode = .{ .immutable = testing.io } });
+    defer database.close();
+    try testing.expectEqual(@as(u32, 1), try readVersion(&database));
+    try testing.expectEqual(@as(u32, 2), target_version);
+
+    // A write through this connection is refused by SQLite, not by convention.
+    try testing.expectError(error.ReadOnly, database.exec("DELETE FROM schema_version;"));
+    try testing.expectEqual(@as(u32, 1), try readVersion(&database));
+}
+
 test "delete_order and content_tables name exactly the tables the schema creates" {
     var database = try openMigrated();
     defer database.close();

src/storage/repositories/clients_repo.zig

@@ -2,8 +2,11 @@
 //!
 //! `listClients` returns only `hand_edited = 1` rows. A client the server
 //! materialised from live traffic is runtime state, not configuration, and must
-//! not appear in an export. `countClients` counts **all** rows, because S5's
-//! "has this database ever been configured" predicate needs the true count.
+//! not appear in an export. `hand_edited` is the only marker of operator intent
+//! in this table, so it also decides what `import.isEmpty` counts: a database
+//! carrying nothing but materialised rows has never been configured, and a seed
+//! file must still be able to fill it. `countClients` counts **all** rows and is
+//! a test helper — it deliberately does not answer that question.
 //!
 //! The import path is list / insert / deleteAll / count, plus the two runtime
 //! calls `upsertSeen` and `pruneStale` that the Phase 7 client tracker owns.
@@ -134,7 +137,8 @@ pub fn deleteAllClients(database: *db.Db) db.Error!void {
     return database.exec("DELETE FROM clients;");
 }
 
-/// Counts every row, including the ones `listClients` filters out.
+/// Counts every row, including the materialised ones `listClients` filters out.
+/// Used by tests; `import.isEmpty` counts operator intent instead.
 pub fn countClients(database: *db.Db) db.Error!i64 {
     return database.queryInt("SELECT count(*) FROM clients");
 }

src/storage/repositories/context.zig

@@ -10,6 +10,8 @@
 
 const std = @import("std");
 
+const safe_url = @import("../../safe_url.zig");
+
 const log = std.log.scoped(.repositories);
 
 /// Group name → `groups.id`, or blocklist source URL → `blocklist_sources.id`.
@@ -17,6 +19,42 @@ pub const IdMap = std.StringHashMapUnmanaged(i64);
 
 const no_ids: IdMap = .empty;
 
+/// The line either lookup writes when the caller's map lacks a name, as a value
+/// rather than a format string at each call site.
+///
+/// Two reasons, in that order. A blocklist source url is operator-supplied and
+/// nothing on the way in stops it carrying a credential in its userinfo, its
+/// path or its query, so it reaches the log through `safe_url.redact` and
+/// through nothing else. And a `std.log` line is not readable from a unit test
+/// under the default test runner, which installs its own `std_options`; the
+/// tests below read this value instead of stderr.
+const MissingId = union(enum) {
+    /// A group name, out of the configuration file or a `groups` row.
+    group: []const u8,
+    /// A blocklist source url, out of the configuration file or a
+    /// `blocklist_sources` row.
+    source: []const u8,
+
+    pub fn format(self: MissingId, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        switch (self) {
+            // A group name holds no credential, so nothing is dropped from it.
+            // It goes through `quoteText` for what that does to any
+            // operator-supplied string: it delimits a name with a space in it,
+            // escapes a `'` that would otherwise close the delimiter, and bounds
+            // a name of any length. The quotes are the type's own, so this
+            // format string adds none.
+            .group => |name| try w.print("no group id for {f}", .{safe_url.quoteText(name)}),
+            // Redaction costs this line the component that told two sources on
+            // one host apart, and there is no row id to name the source by
+            // instead: the id the other call sites print is the one this call
+            // failed to find. The scheme, the host and the port are what is
+            // left. The rule against writing a secret to a log does not bend for
+            // a line that would read better with one.
+            .source => |url| try w.print("no blocklist source id for {f}", .{safe_url.redact(url)}),
+        }
+    }
+};
+
 pub const InsertContext = struct {
     /// Unix epoch seconds, from `std.Io.Clock.real.now(io).toSeconds()`.
     now: i64 = 0,
@@ -29,14 +67,16 @@ pub const InsertContext = struct {
     /// `NotFound` is a member of `db.Error`, so it needs no wider error set.
     pub fn groupId(self: InsertContext, name: []const u8) error{NotFound}!i64 {
         return self.group_ids.get(name) orelse {
-            log.warn("no group id for '{s}'", .{name});
+            log.warn("{f}", .{MissingId{ .group = name }});
             return error.NotFound;
         };
     }
 
+    /// `error.NotFound` means what it means in `groupId`, for the map keyed by
+    /// source url.
     pub fn sourceId(self: InsertContext, url: []const u8) error{NotFound}!i64 {
         return self.source_ids.get(url) orelse {
-            log.warn("no blocklist source id for '{s}'", .{url});
+            log.warn("{f}", .{MissingId{ .source = url }});
             return error.NotFound;
         };
     }
@@ -44,6 +84,48 @@ pub const InsertContext = struct {
 
 const testing = std.testing;
 
+fn expectLine(expected: []const u8, missing: MissingId) !void {
+    var buf: [8 * safe_url.max_len]u8 = undefined;
+    try testing.expectEqualStrings(expected, try std.fmt.bufPrint(&buf, "{f}", .{missing}));
+}
+
+test "a missing source id names where the url points and not what it carries" {
+    // The three components a blocklist url carries a credential in, on the one
+    // line that used to print all three: an api key in the query, userinfo, and
+    // a token in a path segment.
+    try expectLine(
+        "no blocklist source id for https://lists.example",
+        .{ .source = "https://lists.example/hosts.txt?apikey=s3cr3t" },
+    );
+    try expectLine(
+        "no blocklist source id for https://lists.example:8443",
+        .{ .source = "https://user:pa55@lists.example:8443/hosts.txt" },
+    );
+    try expectLine(
+        "no blocklist source id for https://lists.example",
+        .{ .source = "https://lists.example/download/token/hunter2/hosts.txt" },
+    );
+    // What an operator still gets: the scheme, the host and the port.
+    try expectLine(
+        "no blocklist source id for http://10.0.0.2:8080",
+        .{ .source = "http://10.0.0.2:8080/a/hosts.txt" },
+    );
+}
+
+test "a missing group id quotes, escapes and bounds the name" {
+    try expectLine("no group id for 'kids'", .{ .group = "kids" });
+    // A name is database text as well as file text, so a newline in it would end
+    // this line and start one of the operator's choosing.
+    try expectLine(
+        "no group id for 'ads\\n2026-01-01 ERROR forged'",
+        .{ .group = "ads\n2026-01-01 ERROR forged" },
+    );
+    // Nor can a name close the quote around it.
+    try expectLine("no group id for 'kids\\' --'", .{ .group = "kids' --" });
+    const long_name = "n" ** (2 * safe_url.max_len);
+    try expectLine("no group id for '" ++ long_name[0..safe_url.max_len] ++ "...'", .{ .group = long_name });
+}
+
 test "an InsertContext with no maps reports a missing id rather than trapping" {
     const ctx: InsertContext = .{};
     try testing.expectError(error.NotFound, ctx.groupId("default"));

src/tests.zig

@@ -4,6 +4,7 @@ comptime {
     _ = @import("main.zig");
     _ = @import("app.zig");
     _ = @import("version.zig");
+    _ = @import("safe_url.zig");
     _ = @import("dns/types.zig");
     _ = @import("dns/header.zig");
     _ = @import("dns/name.zig");
@@ -31,6 +32,7 @@ comptime {
     _ = @import("storage/db.zig");
     _ = @import("config/model.zig");
     _ = @import("config/validate.zig");
+    _ = @import("config/faults.zig");
     _ = @import("storage/config_schema.zig");
     _ = @import("storage/migrations.zig");
     _ = @import("storage/querylog_schema.zig");

src/upstream/dot_client.zig

@@ -16,11 +16,59 @@ const net = std.Io.net;
 const tls = std.crypto.tls;
 const Certificate = std.crypto.Certificate;
 
+const safe_url = @import("../safe_url.zig");
 const transport = @import("transport.zig");
 const tls_client = @import("../platform/tls_client.zig");
 
 const log = std.log.scoped(.dot_client);
 
+/// One diagnostic line about one client, as a value rather than a format string
+/// repeated at each call site.
+///
+/// Two reasons, in that order. The url is redacted in exactly one place, so a
+/// line added later cannot print it whole — the defect this type closes was four
+/// call sites each formatting `endpoint.url` with `{s}`, missed by three review
+/// rounds because each looked like the three beside it. And a `std.log` line is
+/// not readable from a unit test under the default test runner, which installs
+/// its own `std_options`; the tests below read this value instead of stderr.
+const Diagnostic = struct {
+    endpoint: transport.Endpoint,
+    detail: Detail,
+
+    const Detail = union(enum) {
+        /// `resolveAddress` refused the host.
+        not_an_ip_literal,
+        connect_failed: anyerror,
+        handshake_failed: Handshake,
+        bundle_load_failed: anyerror,
+
+        const Handshake = struct {
+            verify_name: []const u8,
+            cause: anyerror,
+        };
+    };
+
+    pub fn format(self: Diagnostic, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        try w.print("dot upstream {f}: ", .{safe_url.redactQuoted(self.endpoint.url)});
+        switch (self.detail) {
+            // The redacted url ends in the host and the port, so naming the host
+            // again would add nothing but an unredacted copy of it.
+            .not_an_ip_literal => try w.writeAll("host is not an IP literal"),
+            .connect_failed => |err| try w.print("connect failed: {s}", .{@errorName(err)}),
+            // `verify_name` is a host name rather than a url, so it carries no
+            // component redaction could drop. It goes through `quoteText` for
+            // what that does to any operator-supplied string: it delimits it,
+            // escapes it and bounds it.
+            .handshake_failed => |hs| try w.print("TLS handshake as {f} failed: {s} ({t})", .{
+                safe_url.quoteText(hs.verify_name),
+                @errorName(hs.cause),
+                tls_client.classify(hs.cause),
+            }),
+            .bundle_load_failed => |err| try w.print("CA bundle load failed: {s}", .{@errorName(err)}),
+        }
+    }
+};
+
 pub const ResolveError = error{ConnectFailed};
 
 /// DoT endpoints take IP literals. Name resolution for upstreams is out of
@@ -89,6 +137,10 @@ pub const DotClient = struct {
         return .{ .ptr = self, .exchangeFn = exchangeFn };
     }
 
+    fn diagnose(self: *const DotClient, detail: Diagnostic.Detail) Diagnostic {
+        return .{ .endpoint = self.endpoint, .detail = detail };
+    }
+
     fn exchangeFn(
         ptr: *anyopaque,
         io: std.Io,
@@ -119,20 +171,14 @@ pub const DotClient = struct {
         if (query.len > transport.max_message_len) return error.BufferTooSmall;
 
         const address = resolveAddress(self.endpoint) catch |err| {
-            log.warn("dot upstream {s}: host \"{s}\" is not an IP literal", .{
-                self.endpoint.url,
-                self.endpoint.host,
-            });
+            log.warn("{f}", .{self.diagnose(.not_an_ip_literal)});
             return err;
         };
 
         try self.ensureBundle(io);
 
         var stream = address.connect(io, .{ .mode = .stream }) catch |err| {
-            log.debug("dot upstream {s}: connect failed: {s}", .{
-                self.endpoint.url,
-                @errorName(err),
-            });
+            log.debug("{f}", .{self.diagnose(.{ .connect_failed = err })});
             return mapPhase(err, error.ConnectFailed);
         };
         defer closeStream(io, &stream);
@@ -155,12 +201,10 @@ pub const DotClient = struct {
             .stream_write_buffer = self.buffers.stream_write,
         }) catch |err| {
             const cause = concreteHandshake(&tls_stream, err);
-            log.warn("dot upstream {s}: TLS handshake as \"{s}\" failed: {s} ({t})", .{
-                self.endpoint.url,
-                self.verify_name,
-                @errorName(cause),
-                tls_client.classify(cause),
-            });
+            log.warn("{f}", .{self.diagnose(.{ .handshake_failed = .{
+                .verify_name = self.verify_name,
+                .cause = cause,
+            } })});
             return mapPhase(cause, error.TlsFailed);
         };
         defer closeTls(io, &tls_stream);
@@ -214,10 +258,7 @@ pub const DotClient = struct {
         self.bundle.rescan(self.gpa, io, std.Io.Clock.real.now(io)) catch |err| {
             self.bundle.deinit(self.gpa);
             self.bundle.* = .empty;
-            log.warn("dot upstream {s}: CA bundle load failed: {s}", .{
-                self.endpoint.url,
-                @errorName(err),
-            });
+            log.warn("{f}", .{self.diagnose(.{ .bundle_load_failed = err })});
             return mapPhase(err, error.TlsFailed);
         };
     }
@@ -286,6 +327,102 @@ fn receiveFailure(stream: *tls_client.TlsStream, err: anyerror) transport.Exchan
 
 const testing = std.testing;
 
+fn expectDiagnostic(expected: []const u8, url: []const u8, detail: Diagnostic.Detail) !void {
+    var buf: [8 * safe_url.max_len]u8 = undefined;
+    const line = try std.fmt.bufPrint(&buf, "{f}", .{Diagnostic{
+        .endpoint = .{ .scheme = .dot, .url = url, .host = "host.example", .port = 853, .path = "/" },
+        .detail = detail,
+    }});
+    try testing.expectEqualStrings(expected, line);
+}
+
+test "every diagnostic line redacts the url it names the upstream by" {
+    // The endpoints are built by hand rather than parsed, on purpose.
+    // `Endpoint.parse` refuses `@`, `?` and `#` in the authority and refuses a
+    // `.dot` path other than "/", so no url reaching this client through it can
+    // carry a credential in a component `redact` drops. That is a property of a
+    // parser one file away, not of this file, and these lines used to print
+    // whatever `endpoint.url` held. The redaction is what keeps the parser's
+    // rules from being load-bearing here.
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example': host is not an IP literal",
+        "tls://user:hunter2@dns.example/abcd12",
+        .not_an_ip_literal,
+    );
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example:8853': connect failed: ConnectionRefused",
+        "tls://dns.example:8853/abcd12",
+        .{ .connect_failed = error.ConnectionRefused },
+    );
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example': CA bundle load failed: FileNotFound",
+        "tls://dns.example/abcd12?apikey=s3cr3t",
+        .{ .bundle_load_failed = error.FileNotFound },
+    );
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example': TLS handshake as 'one.one.one.one' failed: " ++
+            "CertificateHostMismatch (certificate)",
+        "tls://token@dns.example/abcd12",
+        .{ .handshake_failed = .{
+            .verify_name = "one.one.one.one",
+            .cause = error.CertificateHostMismatch,
+        } },
+    );
+}
+
+test "a diagnostic escapes and bounds the operator-supplied text it prints" {
+    // A control byte in either field would forge a second record on the one
+    // output `std.log`'s own sink does not reach, and an unbounded host or
+    // verification name would write an unbounded line.
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example\\n2026-01-01 ERROR forged': host is not an IP literal",
+        "tls://dns.example\n2026-01-01 ERROR forged",
+        .not_an_ip_literal,
+    );
+    try expectDiagnostic(
+        "dot upstream 'tls://dns.example': TLS handshake as 'a\\nb' failed: TlsAlert (handshake)",
+        "tls://dns.example",
+        .{ .handshake_failed = .{ .verify_name = "a\nb", .cause = error.TlsAlert } },
+    );
+
+    // Doubling every operator-supplied field writes the same line, and each
+    // field is built from the three byte costs at once: a printable byte spends
+    // one character of the budget, `\n` spends two and `\x00` spends four. That
+    // expansion is why `safe_url.max_len` counts printed characters rather than
+    // source bytes, so the bound holds against the widest escape rather than in
+    // spite of it.
+    const long = "h\n\x00" ** (2 * safe_url.max_len);
+    const longer = long ** 2;
+    var short_buf: [16 * safe_url.max_len]u8 = undefined;
+    var long_buf: [16 * safe_url.max_len]u8 = undefined;
+    try testing.expectEqualStrings(
+        try std.fmt.bufPrint(&short_buf, "{f}", .{Diagnostic{
+            .endpoint = .{ .scheme = .dot, .url = "tls://" ++ long, .host = long, .port = 853, .path = "/" },
+            .detail = .{ .handshake_failed = .{ .verify_name = long, .cause = error.TlsAlert } },
+        }}),
+        try std.fmt.bufPrint(&long_buf, "{f}", .{Diagnostic{
+            .endpoint = .{ .scheme = .dot, .url = "tls://" ++ longer, .host = longer, .port = 853, .path = "/" },
+            .detail = .{ .handshake_failed = .{ .verify_name = longer, .cause = error.TlsAlert } },
+        }}),
+    );
+}
+
+test "a diagnostic keeps what a parsed DoT url carries" {
+    // The limit, pinned so it stays visible: a NextDNS DoT upstream is
+    // `tls://abcd12.dns.nextdns.io`, whose hostname is the whole account
+    // identifier. Redaction cannot remove it without leaving no host and an
+    // unactionable line. See `safe_url.SafeUrl`.
+    var buf: [8 * safe_url.max_len]u8 = undefined;
+    const line = try std.fmt.bufPrint(&buf, "{f}", .{Diagnostic{
+        .endpoint = try .parse("tls://abcd12.dns.nextdns.io"),
+        .detail = .not_an_ip_literal,
+    }});
+    try testing.expectEqualStrings(
+        "dot upstream 'tls://abcd12.dns.nextdns.io': host is not an IP literal",
+        line,
+    );
+}
+
 test "resolveAddress accepts IP literals" {
     const v4 = try resolveAddress(try .parse("tls://1.1.1.1:853"));
     try testing.expectEqual(@as(u16, 853), v4.ip4.port);

src/upstream/pool.zig

@@ -40,10 +40,32 @@
 const std = @import("std");
 
 const health = @import("health.zig");
+const safe_url = @import("../safe_url.zig");
 const transport = @import("transport.zig");
 
 const log = std.log.scoped(.upstream);
 
+/// The one line `exchange` writes about a failed attempt, as a value.
+///
+/// It is a value rather than a format string at the call site for the same
+/// reason `dot_client.Diagnostic` is: a `std.log` line is not readable from a
+/// unit test under the default test runner, so the test below reads this
+/// instead of stderr. An upstream url is operator-supplied and a DoH one carries
+/// its credential in the path — `https://dns.nextdns.io/abcd12` is a whole
+/// NextDNS account identifier — so it reaches the line through
+/// `safe_url.redactQuoted`. Quoted rather than bare because the error name
+/// follows it: a redacted authority may still hold a space and a `:`, so
+/// unquoted, a url ending `ok failed: Timeout` would report a failure that did
+/// not happen.
+const AttemptFailure = struct {
+    endpoint: transport.Endpoint,
+    err: transport.ExchangeError,
+
+    pub fn format(self: AttemptFailure, w: *std.Io.Writer) std.Io.Writer.Error!void {
+        try w.print("upstream {f} failed: {t}", .{ safe_url.redactQuoted(self.endpoint.url), self.err });
+    }
+};
+
 pub const Entry = struct {
     endpoint: transport.Endpoint,
     client: transport.Client,
@@ -61,6 +83,11 @@ pub const Entry = struct {
 /// A copy of one entry's health, taken under the mutex. Feeds
 /// `GET /api/upstream/health` in Phase 8.
 pub const Snapshot = struct {
+    /// Whole, not redacted. `GET /api/upstream/health` returns this to a session
+    /// that `GET /api/upstreams` already serves the same url to in full, so
+    /// redacting here would hide nothing from that reader and would make two
+    /// responses of one API disagree. A consumer reachable without a session has
+    /// to redact it itself.
     url: []const u8,
     enabled: bool,
     available: bool,
@@ -165,7 +192,7 @@ pub const Pool = struct {
 
                 const response = result catch |err| switch (transport.group(err)) {
                     .peer_fault => {
-                        log.debug("upstream {s} failed: {t}", .{ entry.endpoint.url, err });
+                        log.debug("{f}", .{AttemptFailure{ .endpoint = entry.endpoint, .err = err }});
                         self.recordFailure(io, entry, completed_at, err);
                         last_fault = err;
                         continue;
@@ -385,6 +412,38 @@ const test_cfg: health.Config = .{
 
 const test_timeout: std.Io.Clock.Duration = .{ .raw = .fromSeconds(10), .clock = .awake };
 
+fn expectFailureLine(expected: []const u8, url: []const u8, err: transport.ExchangeError) !void {
+    var buf: [8 * safe_url.max_len]u8 = undefined;
+    const line = try std.fmt.bufPrint(&buf, "{f}", .{AttemptFailure{
+        .endpoint = try .parse(url),
+        .err = err,
+    }});
+    try testing.expectEqualStrings(expected, line);
+}
+
+test "the failed-attempt line names an upstream by a url carrying no credential" {
+    // A NextDNS DoH upstream puts the whole account identifier in the path, and
+    // this line ran at `debug` on every peer fault, so a debug-level operator
+    // persisted it to the journal once per failure.
+    try expectFailureLine(
+        "upstream 'https://dns.nextdns.io' failed: Timeout",
+        "https://dns.nextdns.io/abcd12",
+        error.Timeout,
+    );
+    try expectFailureLine(
+        "upstream 'https://cdn.example:8443' failed: TlsFailed",
+        "https://cdn.example:8443/d/hunter2/dns-query",
+        error.TlsFailed,
+    );
+    // What it still says, because an operator reading a failover has to know
+    // which upstream failed: the scheme, the host and the port.
+    try expectFailureLine(
+        "upstream 'tls://9.9.9.9:853' failed: ConnectFailed",
+        "tls://9.9.9.9:853",
+        error.ConnectFailed,
+    );
+}
+
 test "Pool satisfies the Client interface" {
     var threaded: std.Io.Threaded = .init(testing.allocator, .{});
     defer threaded.deinit();

src/web/handlers/blocklists.zig

@@ -9,6 +9,10 @@
 //! refresh downloads and compiles before the response is written: 202 is
 //! "accepted and done as far as this connection is concerned", and the status
 //! table in the body is what tells the operator which sources actually landed.
+//!
+//! `DELETE /api/blocklists/{id}` removes the row, reloads, and then sweeps the
+//! compiled files that row named, so `<data_dir>/blocklists/` follows the table
+//! the operator is looking at rather than the scheduler's next pass.
 
 const std = @import("std");
 const Allocator = std.mem.Allocator;
@@ -130,7 +134,27 @@ pub fn applyDelete(state: *server.WebState, io: std.Io, id: i64) ?Failure {
     state.config_lock.unlock(io);
 
     written catch |err| return mutations.dbFailure(err, url_conflict);
-    return mutations.reload(state, io);
+    const failure = mutations.reload(state, io);
+    pruneFiles(state, io);
+    return failure;
+}
+
+/// Removes the compiled files the deleted row leaves behind.
+///
+/// This is the moment an orphan is made during normal operation, and the only
+/// other sweep is the scheduler's — up to `blocklist_update.interval_hours`
+/// away. Without this call a deleted list keeps its megabytes on disk for a day.
+///
+/// After the reload and never part of the response: the row is gone and the
+/// snapshot has stopped enforcing the list, so bytes still on disk are not a
+/// failed delete. `Manager.pruneOrphans` takes the manager's writer lock, which
+/// the reload above has already taken and released — nothing here holds it, and
+/// `state.config_lock` was released before either.
+fn pruneFiles(state: *server.WebState, io: std.Io) void {
+    const manager = state.manager orelse return;
+    manager.pruneOrphans(io) catch |err| {
+        log.warn("pruning the deleted blocklist's files failed: {s}", .{@errorName(err)});
+    };
 }
 
 /// Refreshes every enabled source, then applies the result (ruling 12).

src/web/handlers/mutations.zig

@@ -184,8 +184,12 @@ fn rebuildFailed(what: []const u8, cause: []const u8) Failure {
 const skeleton_group = "default";
 const skeleton_upstream: model.UpstreamServer = .{ .url = "https://dns.example/dns-query" };
 
-/// Runs the shipped validator over `cfg` and returns the first problem's text,
+/// Runs the shipped validator over `cfg` and returns the first failure's text,
 /// or null when the candidate is valid. The text is arena-allocated.
+///
+/// Failures only: a warning describes a configuration that is legal, and a row
+/// this API is about to store cannot be rejected for one. The blocklist source
+/// a POST creates is in no group yet — a warning by design, and never a 400.
 pub fn firstProblem(arena: Allocator, cfg: model.Config) error{OutOfMemory}!?[]const u8 {
     var diags: validate.Diagnostics = .init(arena);
     defer diags.deinit();
@@ -194,8 +198,7 @@ pub fn firstProblem(arena: Allocator, cfg: model.Config) error{OutOfMemory}!?[]c
         error.OutOfMemory => return error.OutOfMemory,
         else => {},
     };
-    if (diags.problems.items.len == 0) return null;
-    const problem = diags.problems.items[0];
+    const problem = diags.firstFailure() orelse return null;
     return try std.fmt.allocPrint(arena, "{s}: {s}", .{ problem.path, problem.message });
 }
 

src/web/metrics.zig

@@ -2,7 +2,9 @@
 //!
 //! Two halves, so that neither needs the other to be testable: `collect` walks
 //! the live collaborators and copies every number into a `Sample`, and `render`
-//! turns a `Sample` into text. Nothing is computed during rendering.
+//! turns a `Sample` into text. Nothing is computed during rendering, with one
+//! exception: an upstream url is redacted where its label is written rather
+//! than where it is copied. `writeUrlLabel` carries the reasoning.
 //!
 //! Three rules the collection half obeys:
 //!
@@ -32,6 +34,7 @@ const logging = @import("../platform/logging.zig");
 const pool_mod = @import("../upstream/pool.zig");
 const rate_limiter = @import("../server/rate_limiter.zig");
 const retention_mod = @import("../storage/retention.zig");
+const safe_url = @import("../safe_url.zig");
 const server = @import("server.zig");
 
 /// The exposition format version, as the 0.0.4 specification writes it.
@@ -96,6 +99,9 @@ pub const DohListenerSample = struct {
 
 /// One upstream, with every string owned by the caller's arena.
 pub const UpstreamSample = struct {
+    /// The configured url, whole. It reaches the exposition only through
+    /// `writeUrlLabel`, which redacts it; a reader of this field is reading a
+    /// credential.
     url: []const u8,
     enabled: bool,
     available: bool,
@@ -377,16 +383,19 @@ fn endpointValue(
 
 fn renderUpstreams(w: *std.Io.Writer, list: []const UpstreamSample) std.Io.Writer.Error!void {
     try labeledHead(w, "nxdns_upstream_up", "1 while an upstream is enabled and healthy.", "gauge");
-    for (list) |entry| try labeledValue(w, "nxdns_upstream_up", entry.url, @intFromBool(entry.available));
+    for (list, 0..) |entry, i| {
+        try labeledValue(w, "nxdns_upstream_up", i, entry.url, @intFromBool(entry.available));
+    }
 
     try labeledHead(w, "nxdns_upstream_enabled", "1 while an upstream is enabled by configuration.", "gauge");
-    for (list) |entry| try labeledValue(w, "nxdns_upstream_enabled", entry.url, @intFromBool(entry.enabled));
+    for (list, 0..) |entry, i| {
+        try labeledValue(w, "nxdns_upstream_enabled", i, entry.url, @intFromBool(entry.enabled));
+    }
 
     try labeledHead(w, "nxdns_upstream_success_rate", "Share of recent exchanges that succeeded.", "gauge");
-    for (list) |entry| {
-        try w.writeAll("nxdns_upstream_success_rate{url=\"");
-        try writeLabelValue(w, entry.url);
-        try w.print("\"}} {d:.4}\n", .{entry.success_rate});
+    for (list, 0..) |entry, i| {
+        try writeUpstreamLabels(w, "nxdns_upstream_success_rate", i, entry.url);
+        try w.print(" {d:.4}\n", .{entry.success_rate});
     }
 
     try labeledHead(
@@ -395,15 +404,19 @@ fn renderUpstreams(w: *std.Io.Writer, list: []const UpstreamSample) std.Io.Write
         "Failures since an upstream last answered.",
         "gauge",
     );
-    for (list) |entry| {
-        try labeledValue(w, "nxdns_upstream_consecutive_failures", entry.url, entry.consecutive_failures);
+    for (list, 0..) |entry, i| {
+        try labeledValue(w, "nxdns_upstream_consecutive_failures", i, entry.url, entry.consecutive_failures);
     }
 
     try labeledHead(w, "nxdns_upstream_successes_total", "Exchanges an upstream answered.", "counter");
-    for (list) |entry| try labeledValue(w, "nxdns_upstream_successes_total", entry.url, entry.total_successes);
+    for (list, 0..) |entry, i| {
+        try labeledValue(w, "nxdns_upstream_successes_total", i, entry.url, entry.total_successes);
+    }
 
     try labeledHead(w, "nxdns_upstream_failures_total", "Exchanges an upstream failed.", "counter");
-    for (list) |entry| try labeledValue(w, "nxdns_upstream_failures_total", entry.url, entry.total_failures);
+    for (list, 0..) |entry, i| {
+        try labeledValue(w, "nxdns_upstream_failures_total", i, entry.url, entry.total_failures);
+    }
 }
 
 /// Every field of a plain counter struct, under one prefix.
@@ -438,12 +451,93 @@ fn labeledHead(
 fn labeledValue(
     w: *std.Io.Writer,
     name: []const u8,
+    index: usize,
     url: []const u8,
     value: u64,
 ) std.Io.Writer.Error!void {
-    try w.print("{s}{{url=\"", .{name});
-    try writeLabelValue(w, url);
-    try w.print("\"}} {d}\n", .{value});
+    try writeUpstreamLabels(w, name, index, url);
+    try w.print(" {d}\n", .{value});
+}
+
+/// The label set every upstream family shares, up to and including the closing
+/// brace. One definition, because six families have to agree on it exactly:
+/// Prometheus identifies a series by its name and its whole label set, so a
+/// family that labelled its samples differently would be a different series.
+///
+/// `index` is the upstream's position in the pool, in the priority order `Pool`
+/// sorts on. It is here because the url alone stopped identifying a series once
+/// it was redacted: two upstreams on one host — the shape a NextDNS account with
+/// two profiles takes — both print `https://dns.nextdns.io`, and two samples of
+/// one name with one label set is a duplicate series a scrape must not contain.
+/// The position is read from the rendered slice rather than carried in
+/// `UpstreamSample`, so no caller can build two samples that claim one index.
+///
+/// What the index is not: a durable key, and the difference is an operator's to
+/// know. `Pool.Snapshot` carries no row id — threading one out of the repository
+/// through the pool to reach here is a larger change than the defect warrants —
+/// so the position is all there is. Removing `upstreams[0]` renumbers every
+/// upstream after it, and one upstream's history then continues under the label
+/// its neighbour used to carry.
+///
+/// What bounds that: the index is only load-bearing when two upstreams share an
+/// origin, which is the case it was added for. Where origins differ, `url`
+/// carries the identity on its own and a reorder moves nothing that a query
+/// grouping on `url` can see. So group on `url`, and read `index` as the
+/// disambiguator between upstreams that group would otherwise merge.
+fn writeUpstreamLabels(
+    w: *std.Io.Writer,
+    name: []const u8,
+    index: usize,
+    url: []const u8,
+) std.Io.Writer.Error!void {
+    try w.print("{s}{{index=\"{d}\",url=\"", .{ name, index });
+    try writeUrlLabel(w, url);
+    try w.writeAll("\"}");
+}
+
+/// The one place an upstream url becomes exposition text.
+///
+/// `/metrics` is `.auth = .open` in `web/routes.zig` and `web.bind` defaults to
+/// `0.0.0.0`, so a url in a label is readable by anything on the LAN without a
+/// session, and a Prometheus that scrapes it keeps that string for as long as it
+/// keeps the series. A NextDNS DoH upstream is `https://dns.nextdns.io/abcd12`,
+/// where the path segment is the whole account identifier, so the label prints
+/// what `safe_url.redact` leaves: the scheme, the host and the port.
+///
+/// The redaction sits here rather than in `collect` because this is where the
+/// open endpoint writes the value. A `Sample` built anywhere else renders
+/// through this function too, so the guarantee cannot be one caller away.
+/// `UpstreamSample.url` stays whole for the same reason it is safe to: nothing
+/// but this function reads it, and the session-authenticated
+/// `GET /api/upstream/health` reports the same pool with the same urls whole.
+///
+/// **`redact` output is not safe to interpolate into a label value, and this
+/// function is the reason it never has to be.** Do not delete the second layer
+/// on the grounds that the first one escapes.
+///
+/// What `redact` gives: it escapes every control character and bounds its
+/// output, so it emits neither a raw newline nor a lone backslash. What it does
+/// not give: it leaves `"` alone. A `"` is legal in the text `redact` keeps, and
+/// it is the one character that ends a label value — so a host holding one
+/// closes the label and lets the rest of the string write label pairs of its
+/// own. That is a defect of this call site, not of `redact`: a `"` needs no
+/// escape in a log line, which is what `redact` was written for.
+///
+/// So `writeLabelValue` runs over the redacted text rather than instead of it,
+/// and the order is the whole point. It also doubles a backslash `redact` wrote,
+/// which is what keeps `\n` in a host from reading as a newline to a parser: the
+/// four bytes `\x1b` arrive at a scrape as `\\x1b`.
+///
+/// A label value's escapes are not a shell's. `quoteText`, which the log lines
+/// use, answers a different question — its delimiter is `'` and it writes its
+/// own quotes — and it is not the tool here.
+fn writeUrlLabel(w: *std.Io.Writer, url: []const u8) std.Io.Writer.Error!void {
+    // `SafeUrl.format` prints at most `max_len` characters, plus the `...` that
+    // marks a truncation. The buffer is that bound, so the write cannot fail.
+    var buf: [safe_url.max_len + 3]u8 = undefined;
+    var redacted: std.Io.Writer = .fixed(&buf);
+    try redacted.print("{f}", .{safe_url.redact(url)});
+    try writeLabelValue(w, redacted.buffered());
 }
 
 /// The three characters the exposition format reserves inside a label value.
@@ -552,22 +646,24 @@ test "a full sample renders the whole exposition, byte for byte" {
     try testing.expect(std.mem.containsAtLeast(u8, text, 1, "nxdns_blocklist_generation 4\n"));
     try testing.expect(std.mem.containsAtLeast(u8, text, 1, "nxdns_disk_free_bytes 1000\n"));
     try testing.expect(std.mem.containsAtLeast(u8, text, 1, "nxdns_disk_sample_failures_total 1\n"));
+    // The label set carries the redaction, so the path of the configured url is
+    // already gone from the golden text.
     try testing.expect(std.mem.containsAtLeast(
         u8,
         text,
         1,
-        "nxdns_upstream_up{url=\"https://dns.example/dns-query\"} 1\n",
+        "nxdns_upstream_up{index=\"0\",url=\"https://dns.example\"} 1\n",
     ));
     try testing.expect(std.mem.containsAtLeast(
         u8,
         text,
         1,
-        "nxdns_upstream_success_rate{url=\"https://dns.example/dns-query\"} 0.9000\n",
+        "nxdns_upstream_success_rate{index=\"0\",url=\"https://dns.example\"} 0.9000\n",
     ));
     try testing.expect(std.mem.endsWith(
         u8,
         text,
-        "nxdns_upstream_failures_total{url=\"https://dns.example/dns-query\"} 1\n",
+        "nxdns_upstream_failures_total{index=\"0\",url=\"https://dns.example\"} 1\n",
     ));
 }
 
@@ -671,15 +767,281 @@ test "cert reload counters render per endpoint, only for the wired stores" {
     try testing.expect(std.mem.containsAtLeast(u8, both, 1, "nxdns_cert_reload_failures_total{endpoint=\"dot\"} 0\n"));
 }
 
+/// Reads a sample line's label set the way a scrape does and returns the number
+/// of label pairs, or null if the line is not one a parser accepts.
+///
+/// A checker rather than a golden string, because the input that exercises it is
+/// a url no parser accepts, and `safe_url.redact` is entitled to change what it
+/// prints for one of those. What may not change is the property: an
+/// operator-supplied byte must not close a label value early, end the line, or
+/// leave an escape sequence behind that means something else to a parser. Only
+/// the format's three escapes are accepted for that reason — a `\x1b` `redact`
+/// wrote reaches here as `\\x1b`, whose backslash is escaped and whose `x1b` is
+/// three ordinary characters.
+fn labelPairs(line: []const u8) ?usize {
+    var i = (std.mem.indexOfScalar(u8, line, '{') orelse return null) + 1;
+    var pairs: usize = 0;
+    while (true) {
+        const eq = std.mem.indexOfScalarPos(u8, line, i, '=') orelse return null;
+        if (eq == i) return null;
+        for (line[i..eq]) |c| if (!std.ascii.isAlphanumeric(c) and c != '_') return null;
+        if (eq + 1 >= line.len or line[eq + 1] != '"') return null;
+
+        i = eq + 2;
+        while (true) {
+            if (i >= line.len) return null;
+            if (line[i] == '"') break;
+            if (line[i] != '\\') {
+                i += 1;
+                continue;
+            }
+            if (i + 1 >= line.len) return null;
+            switch (line[i + 1]) {
+                '\\', '"', 'n' => i += 2,
+                else => return null,
+            }
+        }
+
+        pairs += 1;
+        i += 1;
+        if (i >= line.len) return null;
+        if (line[i] == '}') return pairs;
+        if (line[i] != ',') return null;
+        i += 1;
+    }
+}
+
+/// `name{labels}` — what Prometheus identifies a series by. Null for a sample
+/// that carries no label set.
+fn seriesKey(line: []const u8) ?[]const u8 {
+    const close = std.mem.lastIndexOfScalar(u8, line, '}') orelse return null;
+    return line[0 .. close + 1];
+}
+
 test "a label value escapes the characters the format reserves" {
+    // Every shape an operator-supplied url can take that reaches the label with
+    // a character the format reserves. The assertion is the property, not the
+    // text: these are urls no parser accepts, and `safe_url.redact` may change
+    // what it prints for one of them without changing what this test protects.
+    const hostile = [_][]const u8{
+        "https://a\"b/dns-query",
+        "https://a\nb/dns-query",
+        "https://a\\b/dns-query",
+        "https://a\x1bb/dns-query",
+        "https://user:pa55@h\"ost/dns-query",
+        "https://\"}{=,\"/dns-query",
+        // The shape `safe_url.redact` is being hardened against in this same
+        // wave: a `?` before the last `@`. What it prints is that fix's to
+        // decide; that the label holds it safely is this one's.
+        "https://lists.example?token=prefix@hunter2",
+    };
+
+    for (hostile) |url| {
+        const upstream_list = [_]UpstreamSample{.{
+            .url = url,
+            .enabled = true,
+            .available = false,
+            .consecutive_failures = 2,
+            .total_successes = 0,
+            .total_failures = 2,
+            .success_rate = 0,
+        }};
+        const text = try renderToString(testing.allocator, .{ .upstreams = &upstream_list });
+        defer testing.allocator.free(text);
+
+        // The url wrote no line of its own, and lost none: every line is a
+        // comment or a sample, and the six families contribute six samples.
+        var samples: usize = 0;
+        var lines = std.mem.splitScalar(u8, text, '\n');
+        while (lines.next()) |line| {
+            if (line.len == 0 or std.mem.startsWith(u8, line, "# ")) continue;
+            try testing.expect(std.mem.startsWith(u8, line, "nxdns_"));
+            if (!std.mem.startsWith(u8, line, "nxdns_upstream_")) continue;
+            samples += 1;
+            // Both labels are there, and both values close where they opened.
+            try testing.expectEqual(@as(?usize, 2), labelPairs(line));
+        }
+        try testing.expectEqual(@as(usize, 6), samples);
+    }
+}
+
+test "two upstreams on one host stay two series" {
+    // Redaction costs the url the job of telling two upstreams apart: a NextDNS
+    // account with two profiles is two urls on one host, and both print
+    // `https://dns.nextdns.io`. Two samples of one name with one label set is a
+    // duplicate series, which is a broken scrape rather than a hidden one.
+    const upstream_list = [_]UpstreamSample{
+        .{
+            .url = "https://dns.nextdns.io/abcd12",
+            .enabled = true,
+            .available = true,
+            .consecutive_failures = 0,
+            .total_successes = 5,
+            .total_failures = 0,
+            .success_rate = 1,
+        },
+        .{
+            .url = "https://dns.nextdns.io/efgh34",
+            .enabled = true,
+            .available = false,
+            .consecutive_failures = 3,
+            .total_successes = 9,
+            .total_failures = 3,
+            .success_rate = 0.75,
+        },
+    };
+    const text = try renderToString(testing.allocator, .{ .upstreams = &upstream_list });
+    defer testing.allocator.free(text);
+
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_up{index=\"0\",url=\"https://dns.nextdns.io\"} 1\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_up{index=\"1\",url=\"https://dns.nextdns.io\"} 0\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_successes_total{index=\"1\",url=\"https://dns.nextdns.io\"} 9\n",
+    ));
+
+    // Distinguishable, not merely present. `nxdns_upstream_up` is the family
+    // this matters most in: one of these two upstreams is down and the other is
+    // up, and a reader has to be able to see which. Under one shared label set
+    // the two samples say 1 and 0 of the same series, so a scrape either reports
+    // whichever it read last or rejects the pair — and the down upstream is
+    // invisible either way, on the endpoint an operator watches to find out.
+    var up_keys: [2][]const u8 = undefined;
+    var up_values: [2][]const u8 = undefined;
+    var found: usize = 0;
+    var up_lines = std.mem.splitScalar(u8, text, '\n');
+    while (up_lines.next()) |line| {
+        if (!std.mem.startsWith(u8, line, "nxdns_upstream_up{")) continue;
+        try testing.expect(found < up_keys.len);
+        const key = seriesKey(line).?;
+        up_keys[found] = key;
+        up_values[found] = line[key.len + 1 ..];
+        found += 1;
+    }
+    try testing.expectEqual(@as(usize, 2), found);
+    try testing.expect(!std.mem.eql(u8, up_keys[0], up_keys[1]));
+    try testing.expectEqualStrings("1", up_values[0]);
+    try testing.expectEqualStrings("0", up_values[1]);
+
+    // No two samples in the scrape share a series key, whatever the urls were.
+    var keys: [32][]const u8 = undefined;
+    var count: usize = 0;
+    var lines = std.mem.splitScalar(u8, text, '\n');
+    while (lines.next()) |line| {
+        if (line.len == 0 or std.mem.startsWith(u8, line, "# ")) continue;
+        const key = seriesKey(line) orelse continue;
+        for (keys[0..count]) |seen| try testing.expect(!std.mem.eql(u8, seen, key));
+        keys[count] = key;
+        count += 1;
+    }
+    try testing.expectEqual(@as(usize, 12), count);
+}
+
+test "an upstream url is redacted before it reaches an open endpoint's label" {
+    // `/metrics` is `.auth = .open`, so every label here is readable without a
+    // session by anything that can reach the bind address. A NextDNS DoH
+    // upstream carries the whole account identifier in its path, and a scraper
+    // keeps a label for as long as it keeps the series.
+    const upstream_list = [_]UpstreamSample{
+        .{
+            .url = "https://dns.nextdns.io/abcd12",
+            .enabled = true,
+            .available = true,
+            .consecutive_failures = 0,
+            .total_successes = 3,
+            .total_failures = 0,
+            .success_rate = 1,
+        },
+        .{
+            .url = "https://user:hunter2@dns.example:8443/dns-query?apikey=s3cr3t#frag",
+            .enabled = false,
+            .available = false,
+            .consecutive_failures = 4,
+            .total_successes = 0,
+            .total_failures = 4,
+            .success_rate = 0,
+        },
+    };
+    const text = try renderToString(testing.allocator, .{ .upstreams = &upstream_list });
+    defer testing.allocator.free(text);
+
+    // The four components a credential can live in, none of them exposed.
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "abcd12"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "hunter2"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "s3cr3t"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "frag"));
+    try testing.expect(!std.mem.containsAtLeast(u8, text, 1, "dns-query"));
+
+    // Every family carries the label, so none of the six may keep the whole url.
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_up{index=\"0\",url=\"https://dns.nextdns.io\"} 1\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_enabled{index=\"0\",url=\"https://dns.nextdns.io\"} 1\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_success_rate{index=\"0\",url=\"https://dns.nextdns.io\"} 1.0000\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_successes_total{index=\"0\",url=\"https://dns.nextdns.io\"} 3\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_failures_total{index=\"0\",url=\"https://dns.nextdns.io\"} 0\n",
+    ));
+
+    // The scheme, the host and the port stay: an operator reading a scrape has
+    // to know which upstream a series is about.
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_enabled{index=\"1\",url=\"https://dns.example:8443\"} 0\n",
+    ));
+    try testing.expect(std.mem.containsAtLeast(
+        u8,
+        text,
+        1,
+        "nxdns_upstream_consecutive_failures{index=\"1\",url=\"https://dns.example:8443\"} 4\n",
+    ));
+}
+
+test "a url longer than the redaction bound cannot run past it" {
+    const long_host = "h" ** (4 * safe_url.max_len);
     const upstream_list = [_]UpstreamSample{.{
-        .url = "https://dns.example/a\"b\\c",
+        .url = "https://" ++ long_host ++ "/dns-query",
         .enabled = true,
-        .available = false,
-        .consecutive_failures = 2,
+        .available = true,
+        .consecutive_failures = 0,
         .total_successes = 0,
-        .total_failures = 2,
-        .success_rate = 0,
+        .total_failures = 0,
+        .success_rate = 1,
     }};
     const text = try renderToString(testing.allocator, .{ .upstreams = &upstream_list });
     defer testing.allocator.free(text);
@@ -688,7 +1050,8 @@ test "a label value escapes the characters the format reserves" {
         u8,
         text,
         1,
-        "nxdns_upstream_up{url=\"https://dns.example/a\\\"b\\\\c\"} 0\n",
+        "nxdns_upstream_up{index=\"0\",url=\"" ++
+            ("https://" ++ long_host)[0..safe_url.max_len] ++ "...\"} 1\n",
     ));
 }
 

src/web/web_integration_test.zig

@@ -1129,6 +1129,93 @@ test "W10 a rule mutation reloads the snapshot and the change is live" {
     try bounded(env.io(), default_budget, mutationReloads, .{ env.io(), env });
 }
 
+// ---------------------------------------------------------------------------
+// deleting a source takes its compiled files with it (m13 ruling F-f)
+// ---------------------------------------------------------------------------
+
+/// The id in a `201 Created` body from `/api/blocklists`.
+fn createdId(body: []const u8) !i64 {
+    const marker = "\"id\":";
+    const at = std.mem.indexOf(u8, body, marker) orelse return error.TestNoId;
+    const rest = body[at + marker.len ..];
+    const end = std.mem.indexOfNone(u8, rest, "0123456789") orelse rest.len;
+    return std.fmt.parseInt(i64, rest[0..end], 10);
+}
+
+fn writeCompiled(io: std.Io, dir: std.Io.Dir, id: i64, body: []const u8) !void {
+    var buf: [64]u8 = undefined;
+    try dir.writeFile(io, .{
+        .sub_path = try std.fmt.bufPrint(&buf, "{d}.list", .{id}),
+        .data = body,
+    });
+    try dir.writeFile(io, .{
+        .sub_path = try std.fmt.bufPrint(&buf, "{d}.wild", .{id}),
+        .data = "",
+    });
+}
+
+fn accessCompiled(io: std.Io, dir: std.Io.Dir, id: i64) !void {
+    var buf: [64]u8 = undefined;
+    return dir.access(io, try std.fmt.bufPrint(&buf, "{d}.list", .{id}), .{});
+}
+
+fn deleteSweepsCompiledFiles(io: std.Io, env: *Env) anyerror!void {
+    var conn: Conn = undefined;
+    try conn.connect(io, env.addr);
+    defer conn.close(io);
+
+    var body_buf: [4096]u8 = undefined;
+    try conn.request("POST", "/api/blocklists", null, "{\"url\":\"https://doomed.test/a.txt\",\"name\":\"doomed\"}");
+    var response = try conn.receive(&body_buf);
+    try testing.expectEqual(@as(u16, 201), response.status);
+    const doomed = try createdId(response.body);
+
+    try conn.request("POST", "/api/blocklists", null, "{\"url\":\"https://kept.test/b.txt\",\"name\":\"kept\"}");
+    response = try conn.receive(&body_buf);
+    try testing.expectEqual(@as(u16, 201), response.status);
+    const kept = try createdId(response.body);
+
+    // The files a refresh would have produced for each row. Neither row carries
+    // a checksum, so the reload the delete runs treats both as never fetched
+    // and reads neither — this case is about the directory, not the snapshot.
+    _ = try env.tmp.dir.createDirPathStatus(io, "blocklists", .fromMode(0o700));
+    var dir = try env.tmp.dir.openDir(io, "blocklists", .{ .iterate = true });
+    defer dir.close(io);
+    try writeCompiled(io, dir, doomed, "doomed.example\n");
+    try writeCompiled(io, dir, kept, "kept.example\n");
+
+    var target_buf: [64]u8 = undefined;
+    const target = try std.fmt.bufPrint(&target_buf, "/api/blocklists/{d}", .{doomed});
+    try conn.request("DELETE", target, null, null);
+    response = try conn.receive(&body_buf);
+    try testing.expectEqual(@as(u16, 204), response.status);
+
+    // The row is gone, so its files are orphans; without a sweep on this path
+    // they would sit here until a restart or the scheduler's next pass.
+    var name_buf: [64]u8 = undefined;
+    try testing.expectError(error.FileNotFound, dir.access(
+        io,
+        try std.fmt.bufPrint(&name_buf, "{d}.list", .{doomed}),
+        .{},
+    ));
+    try testing.expectError(error.FileNotFound, dir.access(
+        io,
+        try std.fmt.bufPrint(&name_buf, "{d}.wild", .{doomed}),
+        .{},
+    ));
+    try accessCompiled(io, dir, kept);
+}
+
+test "W10 deleting a blocklist deletes its compiled files and spares the others" {
+    if (!build_options.integration) return error.SkipZigTest;
+
+    const gpa = testing.allocator;
+    var env = try Env.create(gpa, .{});
+    defer env.destroy();
+
+    try bounded(env.io(), default_budget, deleteSweepsCompiledFiles, .{ env.io(), env });
+}
+
 // ---------------------------------------------------------------------------
 // pause via the API changes a real handler decision (ruling 15)
 // ---------------------------------------------------------------------------