fix(ci): eliminate dev test failure baseline

[bloodf]

Summary

This is Stage 2 of the dev-recovery program. It starts from the verified
post-#124 dev, fixes the remaining deterministic regressions, deletes all 51
stale failure-baseline rows, and restores a zero-raw-failure contract before
the normal PR queue resumes.

• Base: dev@6457d95dd5e16e25cdd758d32dad904d16390256
• Head: 9d73875704a2fec0683ea514a2bb40c7a7c06395
• Baseline: 51 entries -> 0 entries; no additions
• Final fail-closed result: raw 0 / known 0 / stale 0

No feature or port PR should merge until this PR is squash-merged and the
post-merge dev checks are green. Nightly remains manually disabled until
that post-merge verification completes.

Authoritative starting failures

GitHub Actions run 29073976751 / artifact 8220013321 on the exact base
reported 17 raw failures, all 17 known, 34 stale rows, and zero additions:

1. Kiro full-body golden request snapshot.
2. AgentRouter Claude passthrough/context contract.
3. AgentRouter's invalid mixed multi-transport assertion.
4. No-auth catalogs hidden by unrelated connections.
5. Command Code validation model/route selection.
6. GitLab Duo environment base-URL fallback.
7. AgentRouter alias registration.
8. AgentRouter local icon registration.
9. Api.airforce validation source-regex assertion.
10. Suggested-model OpenAI filtering.
11. Static registry export coverage.
12. Copilot/M365 specialized executor registration.
13. VeoAI Free video executor routing.
14. Pollinations missing-connection public fallback.
15. Pollinations excluded-connection public fallback.
16. Pollinations keyless validation.
17. ZenMux Free specialized executor registration.

Runtime and CI changes

• Preserve dotted Kiro catalog IDs on the wire and remove only synthetic
  thinking/agentic suffixes.
• Restore AgentRouter's single Claude /v1/messages transport, aliases, icon,
  passthrough behavior, and context contract.
• Give both Command Code IDs an explicit validation model and share its probe
  status policy between validation and saved-connection health checks.
• Restore GitLab Duo base-URL precedence, including the legacy environment
  variable.
• Keep no-auth catalogs visible; implement Pollinations dual-auth behavior,
  prefer real keys, never forward placeholders, and hide premium-only models
  without a usable key.
• Use a negative explicit non-chat filter for OpenAI-compatible model catalogs.
  The fetch route accepts only exact registry-declared URL/filter pairs and
  refuses redirects.
• Register Copilot Web, M365 Copilot, VeoAI Free, ZenMux Free, and supported
  aliases with their specialized executors; regenerate AGENT-INDEX.md.
• Harden Copilot/M365/ZenMux session handling: bounded asynchronous hashcash,
  no token-keyed session pool, strict M365 host/path validation, credential-safe
  diagnostics, and abort propagation for streaming and non-streaming readers.
• Redact credential headers, JSON/body fields, query/userinfo/fragment values,
  response data, stream chunks, and error messages/stacks in request logging.
• Make Nightly build state isolated and keep publication explicitly gated on
  all successful install/lint/index/build/test steps.

Source provenance and intentional divergences

• OmniRoute source refreshed at
  9cd18bf9a11b7d2e8c037c374631b492adabf469.
• Upstream 9router source refreshed at
  b10b8070632cbdeb3c6b26b0ccf60f433ebf5014.
• Dotted Kiro IDs follow live-verified OmniRoute PR #6170 merge b3a2cfe;
  conflicting dash-conversion experiments were rejected.
• PR test(baseline): drain remaining 3-day window regressions (batch 2) #126 was reviewed only as a patch bank at
  d5acb2fa909e4cf0419ef7cef528db312e5ae2d8:
  • retained/improved: no-auth visibility and ZenMux registration;
  • rewritten: AgentRouter assertions, suggested-model filtering,
    Pollinations fallback, and web-session routing;
  • already landed: Api.airforce canonical validateUrl probing;
  • rejected: baseline additions, positive model allowlisting, unused-helper
    edits, generated-index naming workarounds, and mixed AgentRouter transport.

No #126 commit or baseline hunk was cherry-picked.

Verification

All commands used disposable HOME and DATA_DIR and Node 20.20.2 / npm
10.8.2.

Clean Linux pass 1 (fresh container/cache, root + tests npm ci): exit 0
  280 files passed, 12 skipped; 2,556 passed, 18 expected-fail, 59 skipped
  2,633 assertions total
  log sha256 acc4a0639b7b613cc928699de392abba05cdbedbe6b87d1d513aa46e29ec9281

Clean Linux pass 2 (independent fresh container/cache): exit 0
  identical counts
  log sha256 b06f8fecdb8a82719052aceaa5719b716502a4fe22a5d47ddfcf72124dd753aa

Final fail-closed gate: exit 0
  292 suites / 2,633 assertions / failures 0 / runtime errors 0
  raw 0 / known 0 / stale 0 / baseline additions none
  JSON  sha256 9380115571a3e320039a69592ef800c55de6c417711063cd3e412e3b51662422
  JUnit sha256 03fc8ede15d12b5308e0f9367575c0870685e2bcc5087a5faac6048315416cdd

Focused QA: 25 files / 548 tests passed
npm run lint: exit 0 (181 pre-existing warnings, 0 errors)
npm run check:agent-index: exit 0
npm run build: exit 0; all DB paths under the temporary DATA_DIR
commitlint origin/dev..HEAD: exit 0
git diff --check: clean

The full suite still emits non-failing MaxListenersExceededWarning diagnostics
for DB shutdown listeners. This is pre-existing diagnostic debt, not a hidden
test failure; JSON/JUnit report no unhandled or runtime errors.

GitHub Ubuntu x64 run 29077381388 passed twice on fresh runners. Attempt 1
artifact 8221325378 and attempt 2 artifact 8221414954 each contain 292
suites / 2,633 assertions, zero failures/runtime-error suites, JUnit failures
0, raw/known/stale 0, and no baseline additions. Lint/build run 29077381460
and commitlint run 29077381416 also passed on both attempts.

Provider-authorized live smoke tests were not run because this change must not
read operator credentials.

Documentation, migration, and compatibility

• Added docs/development/dev-zero-test-recovery.md and updated recovery,
  provider-port, test, and Nightly documentation.
• Added focused tests for every behavior change, SSRF controls, validation,
  cancellation, and credential redaction.
• No database migration is added.
• Stored API-key secrets are not rewritten, rotated, exposed, or re-keyed.
• Legacy commandcode, GitLab environment, provider-name, header, API-key,
  and data-path wire compatibility is preserved.

Merge checklist

• Empty known-fails.txt; no additions
• Two independent clean Linux installs and direct raw-zero runs
• Final fail-closed JSON/JUnit reports parse and contain zero failures
• Lint, generated index, isolated build, commitlint, and diff hygiene pass
• Staff review: no local blocker
• QA review: no local blocker
• Required GitHub checks green twice on fresh Ubuntu runners
• All review threads resolved (zero threads at final pre-merge refresh)
• Squash-merged as d0663fd60d782a74aa85bbdd5d0959ad41027e68
• Post-merge dev runs 29077864260 and 29077864283 green; artifact
  8221524665 reports raw/known/stale 0 and 2,633 assertions
• Nightly re-enabled; test(baseline): drain remaining 3-day window regressions (batch 2) #126 closed with the patch-bank audit