feat(git): persist per-commit change-risk history

[RaghavChamadiya]

What

Adds a per-commit git_commits table so the indexer stops throwing away
commit-level history. Each row records one commit in the indexed window:

• author, timestamp, subject
• Kamei change features — lines added/deleted, files / directories /
  subsystems touched, churn entropy, and a fix-flag
• a calibrated change-risk score + level (low/moderate/high)

Previously the git layer persisted only per-file aggregates; commit-level
risk was computable live (repowise risk) but never stored. This is the
data foundation a commits / change-risk surface and change-over-time trends
can build on.

How

• The rows are gathered during the existing repo-wide commit-index walk
  via an optional sink, so there is no extra git pass. The walk already
  parses every commit; it now also yields each commit's full file footprint.
• Scoring reuses the existing linear, interpretable change-risk model —
  pure arithmetic on already-parsed diff data, no LLM and no blame at
  runtime. Author experience (the one feature that needs a subprocess in
  the live path) is reconstructed in-memory as a cumulative per-author
  commit count, keeping the pass fully git-free beyond the one walk.
• Wiring is additive and non-breaking: the rows ride on the existing git
  index summary and are upserted next to the file-level metadata. No
  behaviour change for callers that don't read them; empty in
  rename-tracking mode (which uses the per-file walk).

Includes the ORM model, an Alembic migration, and bulk-upsert / delete /
paginated CRUD (sortable by risk or date).

Validation

• New unit + integration tests: the pure feature/risk builder, a CRUD
  round-trip (risk/date sort, pagination, sha-prefix lookup, idempotent
  upsert, delete), and a real-repo run asserting rows are produced with the
  expected shape. Model roster and schema-reconciliation tests updated.
• Measured marginal cost ≈ 14 ms for 273 commits; the collection adds
  nothing measurable to the walk. No re-index required — the table populates
  on the next index.
• ruff check + ruff format clean on all touched files.

Notes

• The persisted change_risk_score is the raw model output. Its ranking
  (review-priority order) is the intended use; absolute level bucketing can
  skew high on repos whose typical commit is large, so a UI surfacing it
  should normalize per-repo rather than show an absolute badge.