…/rationale

A squash merge collapses a feature branch's commits into one commit on the
default branch. WhyGraph already ingests the lost narrative (commit_titles,
comments on the PullRequest row) but dropped both at the final serialization
step. Stage 0 stops dropping them — no schema change.

- mcp/evidence.py: _pr_dict now emits commit_titles + comments (uncapped — the
  evidence tool's consumer is an agent that handles the full lists).
- analyze/rationale_generator.py: _format_pr renders a "Squashed commits" roster
  and a "Discussion" block, clipped by _PrRenderCaps. Caps are threaded purely
  from RationaleConfig via the generator (no get_config reach-in in the
  module-level formatters).
- core/config.py + whygraph.example.toml: three [rationale] rendering caps
  (pr_roster_max_commits=30, pr_discussion_max_comments=20,
  pr_comment_max_chars=500) with the same >=1 validation as max_diff_chars.

Step 2 of the squash-merge recovery plan. Adds an int 0/1 on_default_branch
column to commit (default 1, server_default text("1")) so PR-origin commits
recovered from squash-merged PRs can be flagged 0 and kept out of the
main-walk-only queries (area-history, refactor-walk).

Migration uses a plain op.add_column (native ALTER TABLE ADD COLUMN) rather
than a batch recreate: recreating commit would trip the commit_file_change
foreign key on a populated DB. Additive + server-default backfilled, so
existing rows become on_default_branch=1 and re-scans stay safe.

No new table or pr_id column — the PR<->commit link stays the existing
commit_titles/_linked_prs path (plan 4.3).

Step 3 of the squash-merge recovery plan. When a feature PR is
squash-merged, its original feature-branch commits are fetched once during
the remote scan and persisted as on_default_branch=0 commit rows, linked to
their PR through the existing commit_titles (no link table).

- scan/pr_origin_enricher.py: new PROriginEnricher crawler. Balanced gate
  (plan 3.3/3.5) — a merged PR is enriched when its commit_titles oids are
  absent from commit (squash detection) AND (the merge commit is file-bulk
  OR it collapsed >= pr_origin_min_commits commits). One targeted batched
  git fetch carries only the gated candidates' refs/pull/N/head refspecs
  (never the refs/pull/* wildcard), pinned under refs/whygraph/pull/*.
  Best-effort: a failed fetch or unreadable oid is logged and skipped, never
  failing the scan (plan 6.6). Idempotent across re-scans.
- services/git: GitFetchRefsCmd + GitLogCommitCmd, with thin
  Repository.fetch_refs / Repository.commit_metadata (reuses Commit.from_git_log).
- cli scan: --pr-origins/--no-pr-origins flag (default on), wired as a
  phase-2 sibling of analyze; gated on a resolved GitHub client so it is
  skipped under --no-remote. Added a panel row.
- core/config: AnalyzeConfig.pr_origin_min_commits (default 5) + >=1
  validation + example toml line.
- mcp 4.10 guards: _boring_shas_in and area_history_commits now filter
  on_default_branch == 1 so recovered origin commits never leak into the
  main-walk-only queries (defensive — they carry no commit_file_change rows).

Tests: balanced-gate unit tests, end-to-end work() with a stubbed repo
(asserts only candidate refspecs, on_default_branch=0 rows, dedup of an oid
shared across PRs, graceful fetch-failure degrade), real-git fetch_refs/
commit_metadata tests, config default/validation, and the 4.10 guard tests.
Full suite: 489 passed.

Step 5 (final) of the squash-merge recovery plan. When a queried line
blames to a squash-merged PR's commit that Stage 1 enriched,
collect_evidence now re-blames the same range at the PR's head_sha so each
line maps back to the original feature-branch commit that authored it —
surfaced as source="pr-origin".

- mcp/evidence.py: _attribute_squash_origins re-blames at head_sha, modelled
  on _predecessor_blame (same rev=-blame call, same best-effort per-PR
  GitError swallow). _enriched_squash_prs_for gates on the correct
  §4.8 predicate: blame SHA == a PR's merge_commit_sha AND that PR has >= 1
  on_default_branch=0 origin row — gate-agnostic, so it covers both
  file-bulk and commit-rich enriched squashes. Hunks feed the existing
  labeled_hunks list, so dedupe / priority / cap machinery is unchanged.
- _SOURCE_PRIORITY: pr-origin = 0.5 (just below blame=0); a real authoring
  commit reached through the squash beats every weaker label but loses to a
  direct HEAD blame hit.
- Labels (§4.9): _SOURCE_LABELS entry (rationale_generator); CommitEvidence.
  source docstring enum (analyze/rationale); evidence.py module +
  collect_evidence docstrings updated four→five signals.

Step 4 (lazy diffs for origin commits) needed no code: the enricher leaves
llm_description NULL and origin commits are normal-sized, so
backfill_evidence_descriptions already routes them through the normal /
backfill_all branch, and Repository.diff resolves now the object is local.

Tests: real squash repo (feature branch squashed into main, tip pinned
under refs/whygraph/pull/1, branch deleted) asserting pr-origin entries
match git blame head_sha; small/non-file-bulk squash variant (§4.8);
graceful degrade on a GC'd head_sha; unenriched squash → no attribution;
source-priority unit test. Full suite: 493 passed.