fix: fal drift probe skips on 403/locked/infra instead of poisoning baseline

[jpr5]

The regression

The live fal queue-lifecycle canary merged in #327 (falQueueLifecycleCanary in src/__tests__/drift/providers.ts, driven by the live leg in src/__tests__/drift/fal-queue.drift.ts) submits to the real fal queue. With FAL_KEY set but the fal account balance exhausted, fal returns:

403 {"detail":"User is locked. Reason: Exhausted balance. Add funds to continue."}

The probe threw a hard InfraError on this. The drift collector cannot parse an InfraError into a drift finding, so it routes it to exit-5 quarantine — which makes the "Base drift report" step (run against main) fail on every PR and on the scheduled drift run. Observed:

InfraError: INFRA_ERROR: fal Queue Lifecycle — fal queue submit: API returned 403
(https://queue.fal.run/fal-ai/flux/schnell): {"detail":"User is locked. Reason: Exhausted balance..."}

This is a main-baseline regression: it poisons the drift baseline for every PR, not just fal work. It blocks #325.

The infra-skip fix

Infra-class fal statuses are an availability signal, not drift. The canary now converts 401 / 402 / 403 (incl. user-locked / exhausted-balance) / 429 / 5xx at any lifecycle step (submit, status) into a distinguished FalCanarySkip, which the live leg catches and turns into ctx.skip() with a clear reason:

fal infra/auth unavailable (403 user-locked) at submit — skipping live canary

Key properties:

• FalCanarySkip is not an InfraError, so withInfraErrorTag re-throws it untouched and the collector never quarantines it.
• The submit-step skip fires before the job is enqueued (no cost, nothing to cancel); the status-step skip is rethrown only after the cancel fires, so a job is never left to run.
• Not over-broadened: only infra-class statuses convert. A genuine 2xx-with-wrong-shape drift is never an infra status, so it still flows through parseJsonResponse → triangulate and still fails / reports critical drift. The expected 400 ALREADY_COMPLETED cancel race is untouched.

Change is scoped to the fal probe/leg plus a one-line skip-preserving guard in the shared withInfraErrorTag. Test-infra only — no shipped-bundle change, no version bump.

Red-green proofs

RED→GREEN #1 — the regression (stubbed fal 403 user-locked at submit):

• RED (pre-fix) — ERROR CLASS : InfraError / MESSAGE : INFRA_ERROR: fal Queue Lifecycle — fal queue submit: API returned 403 ... Exhausted balance → quarantine.
• GREEN (post-fix) — ERROR CLASS : FalCanarySkip / STATUS : 403 / MESSAGE : fal infra/auth unavailable (403 user-locked) at submit — skipping live canary.

Guard — real drift must still be caught (stubbed fal 2xx submit whose queue_position is a string where the mock has a number):

• The canary does not skip (200 is not infra); the leg's exact triangulate(exemplar, real, mock) still yields a critical diff → the leg still fails on real drift. Proven green in fal-canary-skip.test.ts.

New regression/guard tests live in src/__tests__/drift/fal-canary-skip.test.ts (3 tests: 403-user-locked → skip, 401/402 → skip, 2xx-wrong-shape → still drifts). Full unit suite (4643 passed / 44 skipped) and full drift suite (93 passed / 65 skipped) green; tsc --noEmit on touched files clean; prettier/eslint/build clean.

Impact

Restores main's drift baseline (the "Base drift report" step no longer quarantines on a locked fal account) and unblocks #325. When FAL_KEY is later funded, the live canary runs for real again; while the account is locked it honestly skips.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@329

commit: 062a650

[jpr5]

CI note — the one red check (drift-live-pr) is expected and clears on merge

drift-live-pr runs two legs. The base leg checks out origin/main and runs the collector against main's still-buggy probe — so on this PR it reproduces the exact regression live:

InfraError: INFRA_ERROR: fal Queue Lifecycle — fal queue submit: API returned 403
  ... User is locked. Reason: Exhausted balance...
WARNING: 1 test failure(s) could not be parsed as drift reports — quarantined for review (exit 5).
##[error]Base collector quarantined unparseable output (exit 5) — manual triage required

The head leg (this PR's fixed code) was skipped only because the base leg fails first. The base leg cannot go green on the PR because it re-runs main by design — that is precisely why this is a main-baseline regression, and it clears the moment this merges (scheduled main runs then produce clean, reusable base reports and future PRs' base legs run the fixed code).

Every check that gates this PR's own code is green: all 15 test matrix jobs (Node 20/22/24 × Python 3.10–3.13), prettier, eslint, commitlint, build-and-push, exports, agui-schema-drift, scan, preview. The failure is the known 403/exhausted-balance quarantine, not a genuine 2xx envelope drift.

[jpr5]

Superseded by fix/fal-probe-in-progress, which folds in this infra-skip (FalCanarySkip + tests) AND adds the IN_PROGRESS lifecycle handling that the funded-fal reruns exposed. #329 alone couldn't merge (its base step ran main's still-broken probe → poisoned drift-live-pr) and doesn't cover the IN_PROGRESS AssertionError. Branch kept for reference; the consolidated fix re-enables the live leg correctly after the descope lands.