fix(ci): reject regressed performance baselines

[justrach]

Closes #230.

The main performance workflow now fails on normal-threshold regressions without overwriting benchmarks/baseline.json. It still records and commits the failed run to benchmark history for diagnosis.

Validation:

• 599 Python tests passed (30 existing warnings)
• focused regression-policy tests: 2 passed
• Ruff passed
• actionlint passed
• baseline/history at the branch base are unchanged

[github-actions]

Performance Regression Report

Runner: ci-pr | Duration: 5s per endpoint | Threads: 4 | Connections: 100 | Workers: 24

Endpoint	req/s	avg latency	p99 latency	threshold	status
GET /health	82,924	0.22ms	1.05ms	80,000	OK
GET /	86,319	0.30ms	1.30ms	80,000	OK
GET /json	80,342	0.24ms	0.83ms	75,000	OK
GET /users/123	82,960	0.26ms	0.77ms	75,000	OK
POST /items	47,008	0.44ms	1.06ms	45,000	OK
GET /status201	77,062	0.24ms	0.98ms	65,000	OK
AVERAGE	76,103			50,000	OK

✅ All endpoints pass regression thresholds