feat: complete pre-edit context evaluation gate

[Helweg]

Summary

Completes the documented acceptance gate for codebase_edit_context from docs/pre-edit-context-design.md. The tool has shipped across OpenCode, MCP, and Pi, but the evaluation gate was unfinished: no budget file, no graph-neighbor recall threshold, no CI wiring, and no documented baseline.

Changes

• Budget gate: adds the minGraphNeighborRecall threshold to EvalBudget, parseBudget, and evaluateBudgetGate. It is only enforced when the metric is present, so datasets without graph-neighbor expectations are unaffected.
• Budget file: benchmarks/budgets/pre-edit.json with thresholds derived from the measured deterministic mock-embedding baseline.
• CI wiring: eval:pre-edit:ci npm script runs in .github/workflows/ci.yml after the eval smoke check, using the existing mock-embeddings server.
• Tests: three new unit tests for the graph-neighbor recall gate (fail below minimum, pass at minimum, skip when metric absent).
• Docs: baseline metrics documented in docs/pre-edit-context-design.md; changelog updated.

Measured baseline (mock embeddings)

Metric	Baseline
Hit@5	1.0000
MRR@10	1.0000
Graph-neighbor recall	1.0000
p95 latency	687 ms
Response tokens avg / p95 / max	358.5 / 399.5 / 404

Validation

• npm run typecheck: pass
• npm run lint: pass
• Budget/schema/metrics/edit-context tests: 48 passed (3 new budget-gate tests)
• Full gate run end-to-end against mock embeddings: Gate: passed=true violations=[]