claude-code-chat-browser: Benchmark regression gate in CI

[clean6378-max-it]

Calendar Day

Wednesday, June 17, 2026 (PR 2 of 2)

Planned Effort

3 story points — sprint item #6 (Low)

Depends on: Wednesday PR 1 (session cache #4) merged — baselines must reflect cached-path performance.

Builds on: Week 2 PR #76 benchmark harness and benchmarks/baselines.json schema.

Problem

The CI benchmarks job runs pytest tests/benchmarks/ --benchmark-only and uploads artifacts, but is labeled "informational" with no threshold. benchmarks/baselines.json has empty groups — regressions in parse/export/search pass silently.

Goal

One merged PR that populates baselines from a post-cache run, adds a +20% regression gate, documents baseline updates, and renames the CI job to signal it is gated.

Scope

Touch points

• benchmarks/baselines.json — populate means (parse small/medium/large, export, search)
• scripts/check_benchmark_regression.py (new) — compare current vs baseline, exit non-zero if >20%
• .github/workflows/ci.yml — regression step after benchmark run; rename job
• Makefile or docs — make update-baselines command
• Unit test for missing-baseline graceful handling (warn, don't fail)

Gate behavior

• Fail if current_mean / baseline_mean > 1.20
• Missing baselines for new benchmark names: warn, return 0
• Gate on ubuntu-latest job only (avoid cross-OS variance)

Acceptance Criteria

• benchmarks/baselines.json populated from post-cache ubuntu run
• CI fails on injected >20% regression; passes when green
• Missing-baseline case warns without failing (tested)
• make update-baselines (or documented equivalent) regenerates baselines
• Job renamed from "(informational)" to gated label
• PR approved by at least 1 reviewer

Verification

cd C:\Users\Jasen\CppAliance\claude-code-chat-browser
.\.venv\Scripts\Activate.ps1
pytest tests/benchmarks/ --benchmark-only
python scripts/check_benchmark_regression.py benchmark-results.json benchmarks/baselines.json

Out of Scope

• Session cache implementation (Wednesday PR 1 — #4)
• New benchmark scenarios beyond existing three bench files