feat(ci): crude test-presence gate (untested-query MVP)

[veksen]

Implements the MVP tracer bullet for the CI⇄MCP verdict loop — Query-Doctor/Site#3496 (parent epic Query-Doctor/Site#3493).

What & why

Query Doctor only analyses SQL that a real-DB test actually runs against Postgres. A PR that changes data-access code with no such test produces no captured query, so CI goes green having never seen it — and that silence reads as "safe." This gate reports the blind spot honestly instead.

In a PR, when data-access code changes but no real-DB (repository/integration) test changes alongside it, the run is flagged: a [!WARNING] banner in the PR comment (why + next step + flagged files + triage hint) and a failed check. It never claims a query is bad — only that it is unverified, flagged conservatively.

How

• src/gate/test-presence.ts — pure diff heuristic. Buckets the PR's changed files into data-access code vs data-layer tests, and returns a verdict only when data-access changed with zero data-layer tests. No capture, no query-to-site mapping. A repository test counts as a data-layer test, so changing a repository and its test together passes; changing the repository alone flags. Removed/unchanged files never trip it.
• src/gate/changed-files.ts — fetches the changed files from the GitHub API using the GITHUB_TOKEN the action already receives (more reliable than git diff under a shallow actions/checkout).
• main.ts — runs the gate independent of the baseline comparison (fires even on a brand-new PR), threads the verdict into ReportContext, and core.setFaileds on it.
• success.md.j2 — renders the v0 verdict payload as a conservative "could not verify" banner.

Heuristic (deliberately conservative)

Default "data-access" = paths matching *repository* / *.repo.* / dal/ / data-access/; "data-layer test" = a test file whose path signals repository/integration/dal. Tuned to under-fire (epic design principle 3) and to match this project's own *.repository.ts convention. Kept as overridable config data so the per-repo policy config (Query-Doctor/Site#3500) can tune it without touching gate logic.

Tests

src/gate/test-presence.test.ts (13) covers every acceptance criterion — flag / pass-with-test / no-change-unaffected / deletion / framing / diff-introduced-only — plus 2 render tests in github.test.ts. All existing reporter tests still pass. typecheck and build clean.

Scope

The versioned/shared verdict contract, failure taxonomy, and policy engine (Query-Doctor/Site#3497 / #3498 / #3500) live in the Site repo and are separate PRs; this ships the inline v0 payload the issue specifies.

🤖 Generated with Claude Code

[github-actions]

Query Doctor Analysis

3 queries analyzed

0 regressed · 0 improved · 0 new · 0 removed

2 pre-existing issues

• SELECT "guests"."id", "guests"."session_id", "guests"."username", "guests"."avatar_path", "guests"."color", "guests"."side", "guests"."audio_recording_path", "guests"."audio_recording_public", "gue...
  index assets(event_id, inserted_at desc)
  cost 31,003,449 → 1,498 (100% reduction)
• SELECT * FROM guest_ip_addresses WHERE ip_address = '127.0.0.1';
  index guest_ip_addresses(ip_address)
  cost 154,402 → 8 (100% reduction)

_{Using assumed statistics (10000000 rows/table). For better results, sync production stats.}

_{More detail → get_ci_run({ runId: "019f3d3e-313a-7b51-a182-3b0eaa7de21e" }) · view run · docs}