fix(test): retire missing candidate janitor definition

[jamesadevine]

Summary

• remove retired candidate Janitor definition 2557 from the compiler-smoke fixture set
• keep release-backed Janitor definition 2548 documented as the weekly coverage lane
• stop secret rotation from targeting the deleted definition
• update the harness tests, Azure YAML, registry, and architecture documentation to the four live candidate children
• stabilize the canary and failure-reporter fixtures with their proven model/configuration

Why

The candidate compiler smoke currently fails before exercising its children because definition 2557 no longer exists. The live candidate definitions on main are 2554, 2555, 2556, and 2558.

Definition 2564 belongs to the still-open custom-safe-output work in #1508 and cannot be used by main until that feature lands. This PR restores the current lane without weakening or changing the weekly release-backed Janitor pipeline.

The first live run of this branch confirmed all four definitions queue successfully, then exposed pre-existing fixture drift: the canary inherited an invalid assignee and both agent-driven fixtures were using the less reliable smoke model. The small hardening changes here match the configuration already proven green by #1508.

Validation

• npm test -- --run src/compiler-smoke-e2e — 188 tests passed
• npm run typecheck
• npm run build:compiler-smoke-e2e
• GitHub PR checks are green
• live candidate compiler smoke: ADO build 628003 passed all four child pipelines

[azure-pipelines]

Azure Pipelines:
2 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[github-actions]

🔍 Rust PR Review

Summary: Looks good — clean housekeeping change, no Rust code modified.

Findings

✅ What Looks Good

• All five occurrences of the janitor fixture have been consistently removed: config.ts, fixtures.ts, runner.ts, stale.ts, index.ts, the four __tests__/ files, REGISTERED.md, README.md, AGENTS.md, and the secret-rotation script are all updated in lockstep. No dangling references observed.
• The duplicate-definition-ID test at config.test.ts:101–70 is a genuine improvement: the old assertion was toThrow(/canary/) (matching a single duplicate group), replaced with a more precise regex that asserts both duplicate groups (/2601 used by \[canary, azure-cli\].*2603 used by \[noop-target, smoke-failure-reporter\]/). More signal, same test boundary.
• The allowedChangedPaths size check correctly goes from 11 to 9 (4 .md + 4 .lock.yml + .gitattributes).
• Secret rotation script removes only 2557; the still-active 2548 (release-backed janitor) is correctly retained.

No bugs, security concerns, or logic issues found. The PR does exactly what the description says.

Generated by Rust PR Reviewer for #1715 · 33.2 AIC · ⌖ 5.64 AIC · ⊞ 6.3K · ◷

[github-actions]

🔍 Rust PR Review

Summary: Looks good — clean, well-scoped maintenance fix with consistent updates across all affected files.

Findings

✅ What Looks Good

• Consistent sweep: All references to definition 2557 / "janitor" in the candidate lane are removed: config.ts, fixtures.ts, azure-pipelines.yml, the PowerShell secret-rotation script, tests, and docs are all updated atomically. No stale references left behind.
• Duplicate-detection test (config.test.ts line ~101–111): The revised test correctly constructs two separate duplicate pairs (2601 used by canary + azure-cli, 2603 used by noop-target + reporter) and matches the new error format with a regex covering both groups. Logic is sound.
• allowedChangedPaths count: Updated from 11 → 9 (4 md + 4 lock + .gitattributes). Arithmetic checks out.
• smoke-failure-reporter.md allowed-labels fix: Changing allowed-labels: [] to the populated list matching labels: is a correctness improvement — the empty list was likely silently misconfigured.
• Model bumps (canary.md, smoke-failure-reporter.md): gpt-5-mini → claude-sonnet-4.6 looks intentional and consistent across both smoke fixtures.
• stale.ts comment generalization: Removing the hardcoded count ("five fixture builds") in favour of the generic "fixture builds" is the right call — it won't go stale again if the set changes.

No bugs, security concerns, or logic issues found.

Generated by Rust PR Reviewer for #1715 · 38.9 AIC · ⌖ 5.65 AIC · ⊞ 6.3K · ◷