fix(drift): route every provider surface to auto-fix via a surface registry

[jpr5]

The hole

extractProviderName matched drift-block titles against ~9 hardcoded
PROVIDER_MAP keys via text.includes(). ~15 additional provider surfaces
(Cohere, Bedrock, Ollama, fal, ElevenLabs, images, video, moderation,
transcription, rerank, vertex-ai, ...) emit valid, fully-parseable
API DRIFT DETECTED: blocks whose PROSE titles are not keys → extractProviderName
returns null → the block is quarantined (exit 5) → fix-drift.yml gates
auto-remediation on exit 2 only, so those surfaces were never auto-fixed
(mislabeled "manual triage"). models.drift.ts is left as-is (deliberate
PROVIDER_MAP-matching canaries).

The design — structural keying via a single-source-of-truth registry

• SURFACE_REGISTRY (src/__tests__/drift/surface-registry.ts), slug-keyed,
  imported by BOTH schema.ts (validate a slug at emit time) and the collector
  (resolve slug → builderFile/builderFunctions). One table, two consumers —
  adding a surface is a single edit both see.
• formatDriftReport(context, diffs, surface?) emits a machine-readable
  Surface: <slug> marker line. An unknown slug throws at emit time.
• Collector resolution:
  • marker present + known slug → exit-2 entry (auto-fixable) — the fix.
  • marker present + UNKNOWN slug → THROW (loud, exit 1), never silent
    quarantine — a future unkeyed surface fails loudly instead of silently
    routing to manual triage.
  • no-marker legacy block → today's provider-label fallback (exit 5 only for
    truly un-attributable output).
• Every currently-unmapped *.drift.ts emitter now passes its slug;
  builderFunctions filled from the real source exports (not invented).
  Bedrock is split per the spec: bedrock-invoke/bedrock-invoke-stream →
  bedrock.ts, bedrock-converse/bedrock-converse-stream →
  bedrock-converse.ts. Vertex AI keys onto the Gemini path (gemini.ts).
• A registry-coverage test asserts every entry resolves to an existing
  builderFile with non-empty builderFunctions (mirrors fix-drift.ts
  validation, so a bad entry fails locally, not in CI).

Base-report reuse fix (own commit — efficiency, not safety)

The collector wrote only { timestamp, entries, quarantine? }, but the reuse
guard reads report.generatedAt + report.conclusion, so reuse always
failed and every PR ran a fresh ~30-min live base. The collector now writes
generatedAt (ISO alias of timestamp, kept for back-compat) and conclusion
(derived from the exit code: 0→clean, 2→critical, 5→quarantine).

Red-green (fixtures only, no live SDK)

RED (original code) — a Cohere/fal/ElevenLabs drift block routes to
quarantine (len 1, exit 5), never an entry:

× "Cohere /v2/chat (non-streaming)" → exit-2 auto-fix entry (EXPECTED to FAIL)
  → expected [ { …(4) } ] to have a length of +0 but got 1
× "fal.ai sync-run (image payload)" → exit-2 auto-fix entry (EXPECTED to FAIL)
  → expected [ { …(4) } ] to have a length of +0 but got 1
× "ElevenLabs /v1/sound-generation 400 error" → exit-2 auto-fix entry (EXPECTED to FAIL)
  → expected [ { …(4) } ] to have a length of +0 but got 1
Tests  3 failed (3)

GREEN (new code, with the Surface: marker) — all three route to an
exit-2 entry, quarantine empty:

✓ src/__tests__/ws5-red-proof.test.ts (3 tests) 3ms
Tests  3 passed (3)

Surfaces flipped 5 → 2: cohere-chat, fal-sync, elevenlabs (plus every
other previously-unmapped surface: rerank, bedrock-*, ollama, images, video,
moderation, transcription, vertex-ai, fal-queue).

Additional in-suite coverage: unknown-slug marker throws (not quarantine);
isBaseReportReusable accepts a written clean report; registry-coverage test.
Full suite: 4637 passed. format:check, lint, tsc (both tsconfigs),
build, commitlint all clean.

🤖 Generated with Claude Code

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@311

commit: 37733e6