fix(release): drain base-reds r7 of v3.8.51 — connection-test operator-disable race, two vitest teardown leaks

[diegosouzapw]

Drains two reds from the release PR #11442 CI (run 36601097731, head fc8e91b).

1. Unit Tests (6/8): connection-test-respects-operator-disable (2 tests)

Cause — a real production race, not DNS. testSingleConnection() reads the connection (cached) before the probe and decided whether to activate on that snapshot after the probe returned. POST /api/providers fires a background testSingleConnection() on create, so in the test (and in production) this interleaving happens:

1. create → background auto-test reads snapshot isActive:false, no marker ("never activated"), starts probing;
2. operator disables (PUT / bulk PATCH stamps operatorDisabledAt);
3. background probe passes → isOperatorDisabled(staleSnapshot) is false → writes isActive:true. recoverKeyHealth() also rebuilt providerSpecificData from the stale copy, which could drop the marker.

The ENOTFOUND proxy.operator-disable.example.com lines are noise: the undici dispatcher falls back to the stubbed fetch. They only change the probe's timing. The test has been intermittent since it landed with #14941. It failed on ed664fc, 0816f62 and 6f665f3, and passed on 4f65ef3 and 4796624, so it is not a #15101/#15064/#15045 regression.

Fix (src/app/api/providers/[id]/test/route.ts): re-read the row uncached right before the write and use it for the operator-disable check (both the skipped and the valid paths) and as the recoverKeyHealth base.

Proof: new test an operator disable during an in-flight probe is not undone when the probe passes holds the probe on a gate, disables the connection mid-probe, then releases it. It was red on the tip (1 !== 0) and is green with the fix. The file passed 4/4 three times, and all 19 neighbouring files that exercise testSingleConnection passed (159/159).

2. Vitest UI: 2 unhandled errors (all 431 files passed)

• Unhandled Rejection: window is not defined (RawJsonPanel.tsx:85), from tests/unit/translator-friendly-raw-json-panel.test.tsx. The suite never unmounted its React roots. Removing the container does not run effect cleanups, so the 600 ms auto-detect debounce stayed armed and fired after jsdom teardown. The component itself is correct because it clears the timer on unmount. Fix: a mountRoot() helper unmounts every root on cleanup. The new test cleanup leaves no debounced auto-detect running after the test was red before the harness fix (the detect fetch was called 3 times after cleanup) and is green after.
• EnvironmentTeardownError: Cannot load partialWithoutDefaults.ts … after the environment was torn down, from tests/unit/ui/playground-studio.test.tsx. The next/dynamic mock started each tab import without awaiting it. A probe showed all 3 imports still pending when afterAll started ([false,false,false] in 3/3 runs). Fix: the imports are tracked and awaited in afterAll. Unhandled-error detection is unchanged.

Both files passed 3/3 (22 tests).

Gates

typecheck:core OK · check:open-sse-typecheck OK (0) · eslint on the changed files OK · check-file-size OK after commit (the route was tightened back under its 1252-line ceiling).

Note: tests/unit/monitoring-health-cached-credential.test.ts › "returns the stale cached summary immediately" failed once locally at 2262 ms against a 2000 ms budget. That is cold-import timing: it passed 3/3 with this change (1.1 to 1.7 s) and does not touch the connection-test route. It is not addressed here.