QVAC-18421 test[skiplog]: add e2e regression tests for Bergamot vocab cache invalidation

[Victor-Rodzko]

🎯 What problem does this PR solve?

• Regression coverage for QVAC-18420 — for bidirectional Bergamot pairs (e.g. fr↔en) the shared vocab file vocab.<pair>.spm was silently re-downloaded on every loadModel call. Root cause: deduplicateModels dropped one of two byte-identical registry entries, then getModelByPath() returned undefined, expectedSize collapsed to 0, and validateCachedFile wiped the cached vocab.
• The bug was caught by users rather than by our test suite.
• Existing unit tests (update-models-dedup, nmtcpp-resolve-vocab) cover the fix in isolation but don't exercise the end-to-end symptom — silent re-download through the live server path.

📝 How does it solve it?

• Adds 2 e2e tests in tests-qvac covering the shared-vocab branch (vocab.<pair>.spm):
  • translation-bergamot-fr-en-cache-reload
  • translation-bergamot-en-fr-cache-reload
• Each test does load → unload (Round 1, warm cache) then load with onProgress → unload (Round 2, must be a pure cache hit).
• Cache-hit detection is platform-agnostic — counts partial-percentage progress events instead of snapshotting ~/.qvac/models via node:fs. A real re-download emits many downloaded < total events; a true cache hit emits at most one final 100% event per file. The test fails if any partial events are seen on Round 2.
• New TranslationBergamotCacheExecutor lives in tests/shared/executors/ and uses dependency: "none" so ResourceManager evicts in-memory models without touching the on-disk cache.
• Desktop-only: skipped on mobile via SkipExecutor. The bug is in server-side Bare code that's bit-identical across platforms, so desktop coverage is the source of truth and we save expensive Device Farm cycles for a regression that can't manifest differently on the device.

🧪 How was it tested?

• Local desktop run local-local-1778595279738: 19/19 tests passed, both new tests green (128 ms / 129 ms).
• Verified via loadModel.ttfb profiler counter (count: 2) that streaming/onProgress was actually wired up on Round 2 — so an empty-events pass reflects a true cache hit, not a silently disabled callback.
• Server logs confirm both rounds fully validate all 4 companion files (model + lex + vocab + metadata) without any re-download.

[simon-iribarren]

The shape, placement, and intent are right — tests-only PR, test[skiplog] tag, tier1+verify both set, no production-code creep. The mobile SkipExecutor comment ("Server-side Bare code path, identical across platforms — desktop coverage is source of truth") is exactly the level of rationale that prevents future drift, and cache-hit detection via onProgress rather than node:fs snapshots is the right call for mobile portability.

The only red CI check is CodeQL / Analyze (python) failing on a github/codeql-action download 429 — this PR has 0 Python files, so it's unambiguously an infra flake; a rerun should clear it.

Three non-blocking nits worth folding in:

1. Round 2 can pass even if onProgress silently stops firing

translation-bergamot-cache-executor.ts asserts touchedKeys.size === 0, but if a future change drops the onProgress wiring on the cache-hit path, round2 is empty and the test returns passed: true with 0 cache-hit notification(s) — the wrong outcome. Suggest a positive lower bound: assert at least N final percentage === 100 events (one per cached companion file: model + lex + vocab + metadata = 4), which directly mirrors your "Server logs confirm both rounds fully validate all 4 companion files" claim in the description.

2. Round 1 has no onProgress callback

There's no positive evidence Round 1 actually exercised the download path. If something later quietly causes Round 1 to be a cache hit (test reordering, a global fixture pre-warms the cache, or someone removes dependency: "none"), Round 2 trivially passes for the wrong reason and we lose the regression coverage. Two options: (a) attach onProgress to Round 1 too and assert round1.length >= round2.length (Round 1 is at least as active as Round 2) — self-validating regardless of cache state at entry; (b) document explicitly that the regression fires regardless of Round 1's cache state.

3. Pattern overlap with TranslationExecutor is fragile

/^translation-(indictrans|bergamot|llm|salamandra|afriquegemma)-/ also matches translation-bergamot-fr-en-cache-reload. The fix (register TranslationBergamotCacheExecutor first + first-match-wins + inline comment) is correct, but the next person to alphabetize the handler list silently breaks it. Cheapest hardening: tighten TranslationExecutor.pattern with a negative lookahead — /^translation-(...)-(?!.*cache-reload)/ — so the dispatchers are mutually exclusive at the regex level.

Micro-nits

• estimatedDurationMs: 180000 is 3 min for a test you measured at 128 ms locally — probably fine for a cold Round 1 on a slow runner, but 30–60s is plenty of slack and won't mask a hang.
• A one-line comment in translation-bergamot-cache-tests.ts explaining why dependency: "none" (so ConsumerBase doesn't pre-warm the cache before the test even starts) would help a future reader not "fix" it.

Solid regression coverage for QVAC-18420. Approving — CodeQL Python flake aside, none of the above would push me away.

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

[kinsta]

Preview deployments for qvac-docs-staging ⚡️

Status	Branch preview	Commit preview
🔁 Deploying...	N/A	N/A

Commit: bacf19bb2516b71fc97eaf55161e5c93187bcfac

Deployment ID: 870eef2c-eb1f-4bda-9f53-5e0d896d3c01

Static site name: qvac-docs-staging-fazwv