test(ingest): prove the Rust classifier and the SQL compiler agree

[JeremyFunk]

What this layer contains

The differential proof that the two evaluators of one registry agree — no production behaviour changes.

• apps/ingest/src/ai_equivalence_fixtures.rs — drives the real row writer (encode_traces) over a branch-covering span set and emits the fixture corpus, plus telemetry.rs's #[cfg(test)] hookup for it.
• packages/domain/src/ai-registry/__fixtures__/equivalence-spans.jsonl (+ README) — the emitted rows.
• equivalence-support.ts, equivalence.clickhouse.e2e.test.ts, hash-alignment.clickhouse.e2e.test.ts, corpus-equivalence.clickhouse.e2e.test.ts — replay those rows through a real ClickHouse and compare the compiled SQL's verdict to the Rust writer's, span for span.

Why it is shaped this way

The fixtures are generated from the row writer rather than hand-written because the thing under test is the writer's canonicalization — first-occurrence-wins duplicate-key resolution, present-but-empty versus absent, AnyValue → String coercion. A hand-written fixture would encode our belief about that behaviour instead of the behaviour itself, and would keep agreeing with the SQL after the writer drifted. fixture_is_reproducible asserts the emitted bytes are stable, so a drifted classifier fails in Rust before the SQL leg runs at all.

The remaining known divergences are enumerated in PINNED_DIVERGENCES, and the suite fails if that set changes in either direction — a silently-fixed divergence is as much a signal as a new one. hash-alignment separately pins that ClickHouse's cityHash64(value) equals the Rust hash, including adversarial non-ASCII shapes and the negative case that multi-argument cityHash64 is a different function. The fixtures module hooks into telemetry.rs rather than its tests child because encode_traces is private there.

Gates run at this level

• cargo test --locked (+ TRACE_CAPTURE_DIR) — 136 lib + 75 bin, 0 failures, including fixture_is_reproducible and fixture_covers_every_branch.
• packages/domain vitest under CLICKHOUSE_E2E=1 — equivalence 6 passed, hash-alignment 5 passed.
• bun typecheck — 37/37 tasks pass.

corpus-equivalence.clickhouse.e2e.test.ts skips without a local trace-capture corpus, by the same on-demand design as the Rust corpus replay.

Dependencies

Requires #434 — the fixtures drive the wired row writer, and the replay inserts into the traces schema that PR adds. Requires #433 for the SQL compiler and #432 for the classifier.

🤖 Generated with Claude Code

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[JeremyFunk]

Superseded by the v2 stack: #451 → #452 → #453. The registry-artifact + SQL-equivalence design this stack implemented was dropped in the write-side plan's v2 revision; vendor rules are now plain Rust with the vendored classification fixture as the CI gate.

[github-actions]

🍁 Maple PR preview

Note

Preview resources were removed when this pull request closed.

Final commit 5fa7589 · View workflow run