Coverage floors, budget/persistence tests, and a paid-response loss fix

[anantham]

Owner-approved coverage work: lock in today's coverage level, then test the code where a bug costs money or data.

Changes

1. Whole-surface coverage floors (CORE-013 amendment)

• verify:coverage-policy now sums coverage/coverage-final.json with istanbul-lib-coverage, the library behind Vitest's own summary; its totals match the All files row exactly. It fails when a global floor is missed.
• Vitest thresholds still take only per-file entries, because perFile: true would apply a global floor to every file. The guard in vitest.config.ts that rejected positive global floors is removed.
• Floors: lines 60, statements 59, functions 59, branches 47, from the measured baseline rounded down. Recorded in docs/infrastructure/COVERAGE-BASELINE.md and a CORE-013 amendment.
• Validator tests cover pass, miss, an empty total, an unknown metric and an out-of-range value.

2. Fix: a failed OpenRouter price lookup discarded paid translations

• calculateCost runs after a paid call and documents that it must not throw. For an uncached OpenRouter model, though, a failed catalog refresh propagated.
• The unguarded translation paths (the OpenAI/OpenRouter adapter, Gemini translate and claudeService) then dropped the finished translation. The failure now falls through to the existing $0 post-hoc path.
• The regression test failed before the fix.

3. Tests where the risk is

• tests/services/ai/cost.test.ts runs the real hasKnownPricing budget-gate logic, which other suites only mock. It fails closed on missing or non-numeric prices and on a failed refresh, and accepts free models. It also covers OpenRouter refresh-once and date-suffix fallback.
• tests/services/db/translationOps.integration.test.ts runs TranslationOps → facade → repository → fake IndexedDB. It checks that:
  • no API key reaches a persisted settingsSnapshot (verified by mutation: the test fails if narrowing is removed);
  • each retranslation gets the next version, with exactly one active;
  • active switching and deletion work;
  • the stableId:// fallback keeps a translation whose URL cannot be resolved.
• The existing date-suffix cost test never exercised the fallback (its model had its own dated entry); it now does.
• Deleted the unused promisifyCursor, batchOperation and debugTransaction from txn.ts. promisifyCursor also awaited before cursor.continue(), which lets IndexedDB auto-commit the transaction.
• Per-file floors: services/ai/cost.ts L95/F100 (was 68% lines, now 100%), services/db/core/txn.ts L95/F100 (21% → 100%) and services/db/operations/translations.ts L75/F100 (22% → 78%).

Verification

Run on this branch after merging current main (#186 and #187), in a Node 22 container:

• tsc: pass
• eslint: 0 errors
• vite build: pass
• verify:integrity: pass
• vitest --coverage: 9,597 passed, 1 failed. The failure is the known Node-22-only indrasNetImageProvider.test.ts case, which passes on Node 24 per the De-sprawl: retire Sutta Studio shims, single Gemini SDK, delete dead code #186 review.
• Coverage: lines 61.05, statements 59.50, functions 59.58, branches 47.92. There are no threshold errors, and the validator passes all floors.

Audit HIGH 2 (commit-time write durability) and MEDIUM 9 (export API-key leak) were already fixed and tested; this PR doesn't touch them.

🤖 Generated with Claude Code

https://claude.ai/code/session_011d1MbSabs9nA9d9w8pxP1W

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
lexicon-forge	Ready	Preview	Sep 24, 2026 6:56pm UTC