feat(bench): add pytest-benchmark suite with make bench target - #15

What

Implements I2 — performance benchmarks: a benchmarks/ directory (excluded from the regular test run via testpaths) covering alignment, end-to-end compare, loop detection (incl. pathological single-tool repeat), recovery metrics, OpenInference parsing at 1000 spans, and report serialization. Run via new make bench (pytest-benchmark, autosaved locally, .benchmarks/ gitignored).

Why

Roadmap WS-I2 and the evidence-based replacement for the descoped A6: measure where the engine actually strains instead of speculatively building streaming ingestion.

How

• Self-contained synthetic trace factories in benchmarks/conftest.py (deterministic; no coupling to test fixtures).
• Sizes parametrized at 100/500/1000 steps to expose scaling behavior.
• make lint/format extended to cover benchmarks/; dev dependency added: pytest-benchmark>=4.0.

Baseline numbers (M2, py3.14) — the headline finding: LCS alignment dominates everything: 100 steps ≈ 16ms → 500 ≈ 390ms → 1000 ≈ 1.6s (quadratic-ish growth), while loop detection (~3ms @1k), metrics (<1ms), and adapter parsing (~17ms @1k spans) are noise by comparison. This hard-confirms the A6 descope rationale and gives us the regression baseline for future aligner work.

Testing

• make bench: 13 benchmarks pass, results autosaved
• Regular suite untouched and green: 300 passed; lint clean

Checklist

• Exactly one logical change in this PR
• make lint passes
• uv run python -m pytest is green (300 passed)
• Docs updated if user-facing — docs sync PR follows this one
• CHANGELOG.md entry added under [Unreleased]
• Conventional title