fix(gate): stop the untested-data-access gate false-positiving on non-SQL text

[veksen]

Goal

The untested-data-access gate should only fail a build when a PR changes a query that no real-DB test exercises. It is fail by default (packages/core/src/ci/policy.ts:50 in Site), so a false positive blocks a merge that has no data-layer test to add. This PR fixes four false positives reported from dogfooding on our own repo: Query-Doctor/Site#3615 and Query-Doctor/Site#3650.

The precision umbrella is Query-Doctor/Site#3503 (capture-based detection), which replaces this diff heuristic outright. Until that lands, these misfires block merges today. This PR is interim work on the heuristic, not a substitute for #3503.

What

Before: a frontend-only PR that touched no database could fail the build.

• Query-Doctor/Site#3811 (a mobile CSS pass) was flagged on apps/app/src/routes/_app/$username/$projectSlug/queries/index.tsx, a React route component with no SQL, ORM or db import.
• Query-Doctor/Site#3649 (our staging-to-prod merge) was flagged on two .tsx components and on two apps/api/scripts/*.sql files that each ship a real-DB spec in the same PR.

After: neither PR produces a verdict. Genuine data-access changes are still flagged exactly as before.

How

Read src/gate/test-presence.ts first. The two commits are separate mechanisms and are worth reviewing in order.

Commit 1 stops non-SQL text reaching the matcher. Every .tsx flag came from one pattern, the raw select … from shape, matching ordinary TypeScript:

• import { Button, Select } from "@query-doctor/ui" puts Select } from in the text. This is a UI component named Select, not the react-query *.queries.ts import that Query-Doctor/Site#3650 suspected.
• export { default as Select } from "./components/select" in a barrel file hits the same shape one keyword over.
• A JSX comment reading "…and select stay one unwrappable unit… must not wrap away from the control it labels" hits it too. Line-based comment stripping missed it twice: the opening line trims to {/*, not /*, and the continuation lines are bare prose with no * decoration. The quer(y|ies) content match hypothesised in Query-Doctor/Site#3615 does not exist in the config and was not involved.

The fix strips /* … */ spans and module imports before matching. The re-export rule keys on export { and export * rather than a bare export prefix, so a declaration like export const q = sql`SELECT id FROM "users"` is still inspected. That boundary is the one real recall risk in this PR and has its own test.

Commit 2 credits a spec that lives in another directory. baseStem stripped .ts, .py, .go and .rb but not .sql, so it compared backfill-personal-teams.sql against the spec's backfill-personal-teams and found no match. That spec reads the script file and runs it against Postgres. Adding .sql widens an already lenient relation, which is the direction the rule documents: a loose match makes the gate under-fire, the safe side.

What I did not fix

Sweeping every non-test file in Site's apps/app/src, apps/blog/src and packages/ui/src as if fully added, 7 of 415 still classify as data access. All 7 hold SQL as content rather than as a query they run: query fixtures, the rendered CREATE INDEX CONCURRENTLY ON recommendation string, planner-node descriptions, demo SQL in a blog component.

No regex separates "SQL text present" from "SQL text executed". The blunt fix Query-Doctor/Site#3650 proposes, excluding frontend .tsx from the declared set, would create genuine false negatives, because Next.js server components query databases from .tsx. This class is evidence for #3503 rather than something to patch here.

Tests

Five tests in src/gate/test-presence.test.ts, each written before its fix and observed failing first. Four pin a reported false positive. The fifth pins the recall boundary of the re-export rule, and was mutation-checked: it fails under a bare export-prefix rule and passes as written.

Full suite 410/410. npm run typecheck clean.

Beyond the unit tests, I replayed both PRs' real changed-file lists from the GitHub API through evaluateTestPresence. On current main the gate fires on 1 file for Query-Doctor/Site#3811 and 4 for Query-Doctor/Site#3649, matching both bug reports exactly. With this branch, both return no verdict.

For recall, I replayed the last 40 merged Site PRs plus those two. Classification drops exactly the three frontend .tsx files and keeps all 13 genuine data-access files, so nothing real stopped being detected.

Refs Query-Doctor/Site#3615, Query-Doctor/Site#3650

[github-actions]

Query Doctor — 6 successful checks

 Cost regression — No query went up more than 5%
 Untested data access — No changed data-access file without a test
 New query — No new queries
 New query with index recommendation — No new query ships an index recommendation
 Schema drift — No schema changes
 High-value nudge — No index or rewrite past the threshold

---

_{More details via MCP → get_ci_run({ runId: "019fc859-3c55-70bc-9a01-d7c8414d8caf" }) · view run · docs
3 queries read against main on assumed statistics of 10,000,000 rows per table. Sync production stats for costs measured against your real data.}