feat(bench): expansion-gate latency micro-bench harness (#741)

[robotrocketscience]

Summary

Public-side replay path for the #741 acceptance bullet. Adds a
standalone latency micro-bench at benchmarks/expansion_gate_bench.py
that runs retrieve() over a labelled JSONL fixture in four cells:
(gate-on/off) × (bfs-on/off). Per cell × label, reports p50/p95
wall-clock plus the count of LaneTelemetry.expansion_gate_skipped_bfs
fires.

Also ships a 10+10 stub fixture at
benchmarks/fixtures/expansion_gate_stub.jsonl for wiring smoke
tests only — hand-authored from in-repo public sources (#741 body,
CHANGELOG, README, source-file identifiers).

What this does NOT do

The 50+50 labelled corpus that the #741 acceptance bullet actually
requires (broad-prompt p95 ≥ 30% improvement; narrow p50 no regression)
is sourced out-of-tree per the gate:lab-corpus convention on #724.
This PR is the harness + smoke fixture; numbers come from a follow-up
run against the real corpus.

Smoke run

uv run python -m benchmarks.expansion_gate_bench \
    --fixture benchmarks/fixtures/expansion_gate_stub.jsonl \
    --out /tmp/eg_smoke/

Output (excerpt):

cell	broad p95 ms	broad skipped/n	narrow p95 ms	narrow skipped/n
gate-on, bfs-on	4.10	10/10	1.40	1/10
gate-on, bfs-off	1.59	0/10	1.07	0/10
gate-off, bfs-on	1.06	0/10	1.38	0/10
gate-off, bfs-off	1.11	0/10	1.02	0/10

Confirms the gate fires on broad prompts (10/10) and the cell wiring
is correct. Sub-millisecond latencies on the 12-belief multihop
corpus are noise-floor.

Test plan

• CI: import-only check on benchmarks.expansion_gate_bench.
• Local: smoke run against the stub fixture produces JSON with 4 cells × 2 labels.
• Follow-up: run against the out-of-tree 50+50 labelled corpus and post numbers on Adaptive expansion-gate: skip BFS/HRR-expensive lanes on broad prompts #741.

Refs #741.

[sourcery-ai]

Sorry @robotrocketscience, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery

[coderabbitai]

Warning

Rate limit exceeded

@robotrocketscience has exceeded the limit for the number of commits that can be reviewed per hour. Please wait 14 minutes and 57 seconds before requesting another review.

You’ve run out of usage credits. Purchase more in the billing tab.

⌛ How to resolve this issue?

After the wait time has elapsed, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.

Please see our FAQ for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 3f91ab66-dbe8-4a28-a956-230abe55f8ff

📥 Commits

Reviewing files that changed from the base of the PR and between fb3fb3c and deee6b8.

📒 Files selected for processing (3)

• benchmarks/expansion_gate_bench.py
• benchmarks/fixtures/expansion_gate_acceptance.jsonl
• benchmarks/fixtures/expansion_gate_stub.jsonl

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/741-bench-harness

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[github-actions]

PR-size soft cap

This PR is over the advisory size threshold:

• 440 changed lines (limit: 200)
• 3 changed files (limit: 3)

Bigger PRs collide with more open work, which under the parallel-session workflow tends to produce repeated attn:merge-conflict cycles (see #602). When practical, split into smaller PRs that each touch a focused surface.

This is advisory only — nothing is blocked. If the size is intentional (large refactor, module removal, generated code), apply the size:override label and this comment will be removed on the next push.

[robotrocketscience]

[claim:review:fermat:2026-05-14T04:42:09Z]

[robotrocketscience]

Rebased onto github/main (3 atomic commits, all signed G: harness → stub fixture → 100-row acceptance fixture). Verified:

What's clean

• Harness imports resolve on main: expansion_gate.ENV_FORCE_EXPANSION/ENV_NO_EXPANSION_GATE (lines 54-55), retrieval.retrieve/last_lane_telemetry (lines 1071, 1409), benchmark.seed_multihop_corpus (line 409). No dangling symbols.
• Cell wiring: 2×2 (gate-on/off) × (bfs-on/off) covers the Adaptive expansion-gate: skip BFS/HRR-expensive lanes on broad prompts #741 acceptance claim shape (gate-on bfs-on vs gate-off bfs-on, broad p95 ≥ 30% improvement; narrow p50 no regression).
• Telemetry sanity-check: gate_skipped_bfs_count is checked per-cell — non-zero in (gate-on, bfs-on) on broad, zero elsewhere. Body's smoke table confirms 10/10 broad firing.
• No collision with feat(retrieval): adaptive expansion-gate for broad prompts (#741) #743 (merged feat for the gate itself) — this PR is the bench-side replay; Refs #741, not Closes #741. Umbrella stays open for the lab-side numbers.
• Atomic commit ordering: harness module → stub fixture → 100-row fixture. Each commit is independently revertable; bisect-friendly.
• Discretion grep on diff vs main: clean. 100-row fixture's prompts are paraphrased natural-language questions about aelfrice public surface ("how does the four-layer retrieval stack...", "why does the project insist on deterministic stdlib gates..."). Header explicitly states "Hand-authored from in-repo public material (README, CHANGELOG, docs/, src/, tests/)".
• All required checks green after rebase (18 SUCCESS, 4 SKIPPED).

Non-blocking nit (worth a follow-up comment, not a fix-before-merge)

The PR body says "The 50+50 labelled corpus that the #741 acceptance bullet actually requires ... is sourced out-of-tree per the gate:lab-corpus convention on #724." But commit 3 ships an in-tree expansion_gate_acceptance.jsonl with exactly the 50+50 shape. Reading the diff + the file header, the intent appears to be a two-tier corpus: in-tree 100-row public-source fixture for CI-runnable acceptance + future lab-side higher-N corpus for the gated bench. That's defensible, but the naming (expansion_gate_acceptance.jsonl vs expansion_gate_stub.jsonl) and the body framing are in tension — a future reader will wonder which corpus #741 actually accepted against. A one-line clarification in the body or a rename of the 100-row file to e.g. expansion_gate_public_100.jsonl would close that gap. Not blocking — the file itself is self-documenting via its header.

Labeling ready-to-merge.

[github-actions]

merge-train: merged deee6b8 → main via FF push.

[robotrocketscience]

[release:review:fermat:2026-05-14T04:50:37Z]