[Tier 1 / Code] DRY and correctness — C1, C2, C3, C4

[brewpirate]

Tracking issue for Tier 1 code-improvement work from the product review. Source: claude/product-review-recommendations-LXL1v plan at /root/.claude/plans/an-you-look-abundant-journal.md.

Suggested labels (not auto-applied): tier-1, code-improvement.

Items

- [ ] C1. Extract shared store base / mixin

~2,234 lines across the four packages/store-*/src/index.ts files have substantial structural duplication: insertRun, updateRun, insertFailure(s), schema-inspection helpers, row mapping, AbortSignal handling. Extract to packages/core/src/store/adapter-base.ts exposing a template-method base class (or a small set of higher-order helpers). Each backend then provides only the dialect-specific SQL and driver wiring.

• Estimated reduction: 40–50%
• Files: all four packages/store-*/src/index.ts

- [ ] C2. Validate driver row outputs at the boundary

Stores cast driver rows as as unknown as Record<string, unknown> then access fields with String()/Number() coercion (e.g., packages/store-sqlite/src/index.ts:466, :517). If a driver shape changes, this silently produces "undefined" strings or NaN. Apply the same parse() discipline used at insert boundaries: define an ArkType flakyPatternRowSchema and parse rows before mapping. The codebase already does this once at packages/store-sqlite/src/index.ts:432 — generalize it.

- [ ] C3. Stop swallowing fatal plugin errors

packages/plugin-bun/src/pending-writes.ts:33 and packages/plugin-bun/src/preload.ts:296,319 catch all errors and log.warn, even fatal ones (missing tables, bad credentials). For the Bun preload, distinguish fatal-on-startup (auth/migration failure → set process.exitCode=1, fail loud) from transient (retry with backoff, then drain-or-fail). Keeps quiet in the happy path; makes silent data loss impossible.

- [ ] C4. Plugin isolation: catch-all around test wrappers

packages/plugin-bun/src/preload.ts:248–269 and packages/plugin-vitest/src/index.ts:42–53 could let an exception in failure-recording (e.g., a malformed stack from a test wrapper) propagate into the host runner. Wrap recording paths in try/catch that logs and continues — plugin failures must never abort the test run being instrumented (also enforced by the plugin-isolated-failure.md rule in .claude/rules/).

Why this tier

DRY / correctness foundation. C1 is the largest single refactor in the review and pays for itself by making every future store-touching change ~4× cheaper. C2–C4 are small, safe correctness fixes that should bundle in one PR.

Verification

• C1: full integration test suite must pass (Postgres + Turso in-memory + Supabase secrets)
• C2: tests for malformed driver shape (mock driver returning unexpected types) → parse error, not silent coerce
• C3: preload test where store throws on first call → process.exitCode === 1
• C4: preload test where recording throws mid-test → host runner continues
• bun test, bun run test:integration (where DBs available), bun run typecheck, bun run check clean