Unclaimed files left in the canonical checkout. Inventory: canonical-cleanup-inventory.md (commit-as-evidence groups). detect-secrets scan clean on 2026-11-02: hits are sha256 pins, git object ids, test sentinels already in tracked tests and unit-file lines.
3.3 KiB
Row 165 (#1662) round 2 review, Dewey (security)
Request: Gitea 29040. Candidate candidate-manifest.sha256
4bcd6a2ef1c52f5e83bc3aa5412cca7ea13d884f03a6fa481a93e4210d28fa5e (24 files), build.patch
cc886c47…9555. I applied build.patch to a fresh detached worktree at 70c9131f0, and
sha256sum -c on the manifest passed. I ran no tests, since Sage gave me no slot for 165
(sage-0951). The run evidence is Darkwing's round 2 slot run: singles 163 files with pass 1419
and fail 0, and mutants 37 of 37 killed. Those logs are in Darkwing's scratch, which I don't read.
The mutant ties below come from BUILD.md's table and from my own reading of each case.
Verdict: approve.
What changed since round 1
I compared each of the 24 files byte by byte against my round 1 tree (89ca3fcf). The seven that
changed are the seven Darkwing named, and no file was added. webui/src/bus.mjs is still
dfbc2289, the channel-proof pin. The public/ +/- digest is still 939ecd28…9444, so mos-0832
stands. No manifest path changed between 70c9131f0 and HEAD; only SESSIONS.md, QUEUE.md and
queue.json moved.
F1: resolved
instance-settings.mjs set now refuses human and tester with invalid-request. The check
comes right after the id() and op check and before writer.read, so it covers both add and
change and nothing is read or written. ID_PATTERN is lowercase-only, so Human and TESTER
already fail id(). In the broker's organization branch, invalid-request maps to
fail("invalid-request"), the same path as the other shape refusals. The test calls refuses()
for each name, and refuses() also asserts the file bytes are unchanged. Without the check, an
add with merge-gate would write, so mutant F1 has to die. The business README states the rule.
The cost Darkwing names is real but not blocking. The human sees invalid-request with the
generic text, because the broker drops the writer's message (my N3). Choosing invalid-request
over a new code keeps the channel-proof pin where it is, and I accept that trade.
F2: resolved
There are three new cases in webui/tests/instances.test.mjs. Each one has a single part wrong
and the other four parts of the serve.mjs:314 check good:
| Case | Only wrong part | Check that catches it | Mutant |
|---|---|---|---|
revision: 'x' |
revision | the 64-hex regex | F2a |
restart: 'yes' |
restart | typeof … !== 'boolean' |
F2b |
instances: {} |
instances[instance] |
!result.instances[instance] |
F2c |
With any one check gone, its case reaches the 200 path and fails the 503 assert, so each mutant
is tied to exactly one case. The two GET cases (an answer without instances, and one without
definitions, both 502) cover my optional note at serve.mjs:91, with mutants E1 and E2.
Filbert note 2 wording
The organization-definition-refused text changed in both tables (CLI REASONS and Console
INSTANCE_REASONS). The CLI test pins it, and so does a real-broker case in the route test.
Neither text names a variable value or a file path.
Other checks
- Grep for 64-hex pins in tests next to a round 2 file: no hits.
- Follow-ups listed in BUILD.md "Round 2": N1 (stale revision first), N2 (a new instance key can
take a held seat name), N3 and Filbert note 1 (the broker drops the writer's message). Moving
the human/tester rule into
validateBusinessDocumentis left for Sage to place.