fix(db): stop v11 from clobbering non-claude harness labels (#19)

[senna-lang]

Closes #19

Root cause

v11 harness mislabeling (checklist item 1). _migrate_v11_add_canonical_sessions unconditionally ran harness = COALESCE(harness, 'claude') for every pre-existing conversations row, regardless of the actual harness. The very next step, _backfill_exchange_provenance (which infers harness from source_path — opencode.db#, rollout-, /.omp/, /.grok/), only touches rows WHERE harness IS NULL. Since v11 always filled the column first, that guard never matched, so any grok/omp-pi/opencode/codex conversation indexed before v11 stayed permanently mislabeled as claude.

file_renames quotepath (checklist item 3). _run_git_follow called git log --follow --name-status ... without -c core.quotepath=false. Git's default core.quotepath=true renders non-ASCII paths as octal-escaped, double-quoted strings (e.g. a literal "caf\303\251.py"), so resolve_aliases returned that mangled string instead of the real filename — rename tracking silently broke for any non-ASCII path.

grok/omp-pi/opencode git_branch (checklist item 2) — investigated, not a fabrication candidate. Checked each harness's real raw schema against tests/fixtures/harness_logs/README.md, which documents key names/nesting verified directly against real logs (grok: 39 sessions, omp-pi: 99 sessions, opencode: 1 real opencode.db, re-verified 2026-08-30 specifically for adapter implementation):

• grok: ACP session/update envelope (timestamp, method, params: {sessionId, update}) — no git field anywhere in tool_call / tool_call_update / message-chunk entries.
• omp-pi: session envelope is {type: "session", version, id, timestamp, cwd} — only a working directory, no branch.
• opencode: project (id, worktree, vcs, name) / session (id, project_id, directory, time_created, model) / message / part tables — no branch column or key anywhere in the data JSON blobs.

None of the three carry a git-branch field anywhere in their data, unlike claude's gitBranch or codex's session_meta.git.branch. I did not add a live git branch --show-current shell-out against the recorded cwd/worktree as a substitute: that would record the branch at index time, not at conversation time — a different (and less accurate) semantic than what git_branch means for claude/codex, and it risks recording the wrong branch if the user has since switched. Left git_branch=None for all three, with comments at each call site recording this finding so it isn't silently rediscovered later.

Fix

• Extracted the source_path -> (harness, source_session_id) heuristic into a single _infer_harness_and_source_session_id helper in src/codeatrium/db.py, used by both _migrate_v11_add_canonical_sessions and _backfill_exchange_provenance. They now always agree on harness and derive identical session_id hashes instead of one silently overriding the other.
• src/codeatrium/file_renames.py: _run_git_follow now passes -c core.quotepath=false to git log.
• src/codeatrium/indexer.py: documented the git_branch investigation at the three git_branch=None sites (parse_grok_exchanges, parse_omp_pi_exchanges, parse_opencode_exchanges).

Verification

• RED: added test_migration_v11_infers_harness_from_source_path_for_non_claude_sessions (builds a pre-v11 DB with a /.grok/ conversation, asserts init_db doesn't mislabel it claude) — failed with assert 'claude' == 'grok' before the fix.
• RED: added test_resolve_aliases_handles_non_ascii_renamed_path (renames a café.py file in a real git repo) — failed with the octal-escaped path string instead of café.py before the fix.
• GREEN: both tests pass after the fix; kept as regression tests.
• make check (ruff + pyright + full pytest suite) passes clean: 663 passed, 0 lint issues, 0 typecheck errors.

Checklist from the issue:

• v11 leaves non-claude harness inference to the same heuristic as backfill (shared helper, no longer overridden)
• investigated grok/omp-pi/opencode branch capture — none carry the data; documented instead of fabricated
• file_renames git calls use -c core.quotepath=false

[senna-lang]

Priority-1 gap: this only fixes the v11 migration going forward for databases whose PRAGMA user_version is still below 11. Databases that already ran the buggy v11 migration already have harness='claude' and a claude:-derived session ID written; _run_migrations skips v11 for them (already applied), and _backfill_exchange_provenance (~line 422) only selects rows where session_id IS NULL OR harness IS NULL" — which is none of them, since v11 already filled both. So every legacy grok/omp-pi/opencode/codex exchange that was mislabeled by the old buggy v11 stays permanently mislabeled after this fix. Please add an idempotent repair migration/backfill that identifies non-Claude source_path values among already-claude`-labeled legacy rows and rewrites their harness/session_id/canonical exchange id — not just fix the forward path.

[senna-lang]

Good catch — fixed in 328f40a.

Added migration v13 (_migrate_v13_repair_legacy_claude_mislabel) that specifically targets DBs already past the buggy v11 (PRAGMA user_version >= 11). It walks every conversations row, re-derives the correct harness from source_path with the same _infer_harness_and_source_session_id heuristic used by v11/backfill, and for any exchange whose stored harness disagrees, rewrites harness, session_id, source_session_id, and — when the column already exists (i.e. _backfill_canonical_exchange_ids already ran against the stale values) — canonical_exchange_id. It also drops the stale claude-hashed sessions row for that conversation once nothing references it anymore (safe: that id is sha256("claude:" + source_path), unique per conversation since source_path is unique).

It's registered in _MIGRATIONS so it runs exactly once per DB via the normal PRAGMA user_version gate, and is independently idempotent (re-deriving already-correct rows is a no-op) if ever invoked again.

Regression test: test_migration_v13_repairs_exchanges_already_mislabeled_by_old_buggy_v11 builds a DB frozen at user_version=12 with harness='claude' and a claude:-derived session_id/canonical_exchange_id already persisted for a grok conversation — reproducing exactly the state a DB that already ran the old buggy v11 would be in. I verified it fails (assert 'claude' == 'grok') with v13 unregistered and passes with it registered.

Also had to fix test_migration_v12_adds_exchange_conversation_ply_index_to_existing_db, which hardcoded len(_MIGRATIONS) - 1 assuming v12 was the last migration; it now targets _MIGRATIONS.index(_migrate_v12_add_exchange_conversation_ply_index) so it still isolates v12 regardless of how many migrations exist.

make check run directly (not via git hooks, per issue #46): 664 passed, 0 lint issues, 0 typecheck errors.