fix(test): unbreak main - a shared capability doc may name no bound #1052

Merged

coilysiren merged 1 commit from coilyco-ops/fix-red-main into main

2026-08-19 04:46:40 +00:00

AGit

coilyco-ops commented

2026-08-19 04:44:30 +00:00

Member

closes #1053 - main went red 20 minutes ago and nothing stopped; the source changes stay and the test moves

main is red at 446170e, and #1018 inherits that rather than causing it. Both of its test failures reproduce on main with none of its changes.

Bisected: 434063f green, then 2b4305e breaks Deep's prompt budget and 1f6e246 breaks the capability-doc check. Both were pushed straight to main, no PR. CI ran on the tip and reported it correctly - ci / test red after 44s - and the red was left standing through five subsequent merges including mine.

Both source changes are right and stay here. 1f6e246 removed the tool-round and model-call figures from coilyco-general because that copy loads on the Dowel lane and on the domain-neutral HTTP profile, whose deployments set different ceilings, and Dowel was observed live saying 6 rounds and 9 calls against a deployed 12 and 2 (#978). Restoring the sentences to satisfy the test would restore a confirmed live misstatement.

So the test moves instead. A copy that names the bounds still has to name the table's. A copy that says its ceilings are deployment-set is failed for naming a figure anywhere, because the loophole in naming none is naming one somewhere else. Proven by hand in both directions against the real file: smuggling 6 tool rounds back in fails it, and deleting the deployment-set declaration without restoring the figures fails it the other way.

Deep's budget ratchets 12852 to 14336 for the bounded-refusal doctrine 2b4305e added. Its neighbouring comment pointed at the raise ledger #1047 deleted, so it now names what replaced it, and the rule itself is written down in docs/sirens-echo-prompt.md.

just gate green.

Not addressed here, and worth its own decision: a red main stopped nothing. The next PR's author cannot tell an inherited failure from one they caused, which is what this cost to find.

closes #1053 - main went red 20 minutes ago and nothing stopped; the source changes stay and the test moves `main` is red at 446170e, and #1018 inherits that rather than causing it. Both of its test failures reproduce on main with none of its changes. Bisected: `434063f` green, then `2b4305e` breaks Deep's prompt budget and `1f6e246` breaks the capability-doc check. Both were pushed **straight to main**, no PR. CI ran on the tip and reported it correctly - `ci / test` red after 44s - and the red was left standing through five subsequent merges including mine. **Both source changes are right and stay here.** `1f6e246` removed the tool-round and model-call figures from `coilyco-general` because that copy loads on the Dowel lane and on the domain-neutral HTTP profile, whose deployments set different ceilings, and Dowel was observed live saying 6 rounds and 9 calls against a deployed 12 and 2 (#978). Restoring the sentences to satisfy the test would restore a confirmed live misstatement. So the test moves instead. A copy that names the bounds still has to name the table's. A copy that says its ceilings are deployment-set is failed for naming a figure anywhere, **because the loophole in naming none is naming one somewhere else**. Proven by hand in both directions against the real file: smuggling `6 tool rounds` back in fails it, and deleting the deployment-set declaration without restoring the figures fails it the other way. Deep's budget ratchets 12852 to 14336 for the bounded-refusal doctrine `2b4305e` added. Its neighbouring comment pointed at the raise ledger #1047 deleted, so it now names what replaced it, and the rule itself is written down in `docs/sirens-echo-prompt.md`. `just gate` green. Not addressed here, and worth its own decision: **a red main stopped nothing.** The next PR's author cannot tell an inherited failure from one they caused, which is what this cost to find.

coilyco-ops added 1 commit

2026-08-19 04:44:30 +00:00

fix(test): let a shared capability doc name no bound, and ratchet Deep

ci / image-build (pull_request) Successful in 40s

Details

ci / test (pull_request) Successful in 1m11s

Details

ci / publish-echo-image (pull_request) Has been skipped

Details

ci / publish-observed (pull_request) Has been skipped

Details

a3d0ac79d8

main is red, and #1018 inherits it rather than causing it. Two commits
landed straight on main at 04:23 and 04:24 without passing CI: 2b4305e
grew Deep's prompt past its budget, 1f6e246 removed the two sentences
TestTheCapabilityDocsFollowTheHarnessBounds requires.

Both source changes are right and stay. 1f6e246 removed the figures
because coilyco-general loads on the Dowel lane and on the HTTP profile,
whose deployments set different ceilings, and Dowel was observed live
saying 6 rounds and 9 calls against a deployed 12 and 2. Restoring the
sentences would restore a confirmed misstatement, so the test moves
instead: a copy that names the bounds still has to name the table's, and
a copy that says they are deployment-set is failed for naming a figure
anywhere. Both directions are proven by hand against the real file.

Deep's budget goes 12852 to 14336 for the bounded-refusal doctrine
2b4305e added. Its neighbouring comment pointed at the raise ledger #1047
deleted, so it now says what replaced it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>