fix(workspace): route Postgres-backed watch/update through the shared DB instead of local wiki.db

[Aman-goel-04]

Summary

• Fixes the incremental-update gate, the config-drift branch, reconcile_repo_head_commit, and _merged_repo_excludes in update.py. All four assumed a repo-local .repowise/wiki.db as the "already indexed" signal, which never exists under a shared Postgres database, so repowise watch --workspace full-reindexed on every single file change instead of updating incrementally.
• Fixes reconcile_repo_head_commit's early return specifically: on Postgres, repositories.head_commit/updated_at were never stamped after an update, which left the UI's "index behind checkout" warning and freshness timestamp permanently stuck even after a successful sync.
• Gives create_engine() a short_lived parameter (default True) so PostgreSQL uses NullPool for every short-lived engine (one connection per checkout, closed on dispose) instead of SQLAlchemy's default AsyncAdaptedQueuePool (up to 15 idle server slots per engine, and the failure class behind [Bug] Async PostgreSQL cleanup races with event-loop shutdown #2062, a pooled connection surviving past its creating event loop and getting handed to a later, unrelated caller). The two long-lived engines (FastAPI app, MCP server) opt out with short_lived=False to keep pooled reuse across requests.

Related Issues

Fixes #2148

Test Plan

• Tests pass (pytest), full suite (tests/unit/, tests/providers/): 17528 passed, 4 pre-existing failures unrelated to this change (confirmed identical against a clean main via git stash: an OS-path baseline test, a Python-version-dependent regex test, and a test hardcoding a master branch name). tests/unit/workspace/test_incremental_update.py + tests/unit/server/mcp/test_config.py: 19/19 passed.
• Lint passes (ruff check .)
• Web build passes (npm run build) (if frontend changes) — not applicable, Python-only change

Manual reproduction (before fix): two disposable repos against a Docker postgres:16 instance via REPOWISE_DB_URL. Confirmed Postgres held a repositories row for test-repo with a matching state.json.last_sync_commit, and no local wiki.db. Running repowise watch --workspace -v and editing a tracked file reliably produced tier=full on every single-line change across dozens of edits.

Regression tests added:

• test_shared_db_indexed_repo_takes_incremental_path: a repo indexed in a configured shared DB (simulated with a file-backed sqlite+aiosqlite URL, no live Postgres required) takes the incremental path even with no local wiki.db. Fails on main with full pipeline must not run for indexed repos.
• test_shared_db_config_drift_does_not_crash: combines config_changed=True with the shared-DB/no-local-wiki.db case. Regression test for an ordering bug caught during self-review: an earlier version of this fix computed has_persisted_index after the config-drift branch already read it, which would raise UnboundLocalError the moment a shared-DB repo's config actually drifted. The existing test_config_drift_runs_full_reindex only covers the local-SQLite case and never would have caught it.
• test_mcp_lifespan_uses_cli_database_env_var updated to assert the MCP server's engine is created with short_lived=False, locking in the long-lived opt-out so a future cleanup pass can't silently drop it.

On the connection-leak half of #2148: I was not able to reproduce an accumulating leak (a rising floor between update cycles) from watch alone, across ~65 cycles spanning sequential single-repo, sequential dual-repo, and forced-concurrent dual-repo edits against a live Postgres instance with serve + mcp also running. Every observed connection spike drained back to baseline. This is consistent with full_index.py's existing try/finally: await engine.dispose() already disposing correctly on both the success and exception paths. The short_lived/NullPool change in this PR closes the mechanism regardless (an idle pooled connection can never be picked up by a later, unrelated caller after this change, since there's no pool to hold it), rather than relying on a leak having been empirically observed.

Checklist

• My code follows the project's code style
• I have added tests for new functionality
• All existing tests still pass
• I have updated documentation if needed. Docstrings on create_engine, reconcile_repo_head_commit, and _merged_repo_excludes updated to describe the shared-DB behavior

[repowise-bot]

🔍 2 things to check

1. create_engine in .../persistence/database.py signature changed. Review 124 callers outside this PR: .../cli/_repo_session.py::open_repo_db, .../augment_cmd/search.py::_pagerank_file_order, .../augment_cmd/search.py::_search_enrich (+121 more)
2. Run tests/integration/test_mcp.py, tests/integration/test_persistence.py, .../persistence/test_embed_batch_pg.py (+7 more) first: they import the changed files

✅ Health of changed files: 3.3 → 3.5 (+0.1)
_{⚠️ Change risk: moderate, riskier than 52% of this repo's commits.}

✅ Health gate: passed

👀 Suggested reviewers @RaghavChamadiya

---

📊 See the full report for this PR
_{Your repo map with this PR's blast radius lit up, every caller of the contracts it changes, and health before and after. No sign-in. · ⭐ Star Repowise · 📥 Install bot · Silence on a single PR with [skip repowise] in the title · Per-repo toggle on repowise.dev/settings?tab=bot · Updated 2026-09-08 06:50 UTC}