# Handoff Brief — Phases 2-5 of the 2026-08-23 Code Review

**Repo:** `C:\Github\transcription` · **Branch:** `traumatized` · **Baseline commit:** `de18c2e`
**Source of truth:** `docs/reviews/2026-08-23-code-review.md` (§9 Prioritized Action Plan)

Phase 1 is **done and committed**. This brief covers everything after it.

> **Status note.** This is a dated, non-canonical artifact, like everything under
> `docs/reviews/**`. It records a plan, not a contract. Where it disagrees with
> `.github/instructions/**` or the canonical docs, **they win**.

---

## 1. Environment — read this first

- `uv` project on **Windows / PowerShell**. `ruff`, `ty`, and `pytest` are **not on PATH**. Always prefix with `uv run`.
- PowerShell has **no heredoc**. Don't write `python - <<'PY'`. Use `python -c "..."` or pipe a single-quoted here-string (`@'` … `'@ | python -`).
- `&&` only chains *external* commands in PowerShell. Use `;` before PowerShell keywords.
- Ruff config (`ruff.toml`): line length **120**, `force-single-line = true` — **one import per line**. Never combine imports.
- `# noqa: PLR0915` / `PLR1702` is established repo convention; don't strip existing ones.

## 2. Mandatory reading before editing `src/transcription/**`

The repo's instruction table requires these before source edits. They are contracts, not suggestions:

- `.github/instructions/services.instructions.md` — transaction boundaries, model ownership, no service-to-service imports
- `.github/instructions/error-handling.instructions.md` — error categories, retriability, user-safe messaging
- `.github/instructions/ui.instructions.md` — page/component boundaries
- `.github/instructions/documentation-sync.instructions.md` — **docs must be updated in the same change** when contracts or behavior change

## 3. Verification commands

```powershell
uv run ruff check .                    # must be clean
uv run pytest -q -m "not external"     # must be 381+ passing
uv run ty check                        # baseline is exactly 10 diagnostics
```

**The `ty` baseline is 10, and all 10 are false positives** — SQLModel/SQLAlchemy column
descriptors typed as `UUID`/`datetime`/`bool`, so `.is_()`, `.asc()`, `func.count()`, and
`group_by()` appear invalid. They are in `services/photos.py` (8) and
`tests/test_storage_reconciliation.py` (2). **Do not "fix" these by changing code.** Handling
them is task P2-1 below, and the fix is suppression comments, not code edits.

Hard-won gotcha: `typing.Mapping` trips ruff's `deprecated-import`, and `collections.abc.Mapping`
doesn't satisfy `ty` for SQLAlchemy row results. `Sequence[RowMapping]` + `RowMapping` is the
only spelling that satisfies both. Don't rediscover this.

---

## 4. What Phase 1 changed (context you need)

Commit `de18c2e`. Four things, all with tests:

1. **`workflows.py` commit boundary.** `process_queued_job` now commits every page except the
   last individually, then defers the final page's write into `_finalize_batch_outcome` so it
   shares the terminal-status transaction.
   - **`_finalize_batch_outcome` gained a `final_page` kwarg.** If you touch this function, that
     parameter is load-bearing.
   - **Two invariants are in tension here — preserve both.** Intermediate pages must stay
     individually durable (guarded by
     `test_workflows_reliability.py::...::test_transcribed_page_is_committed_before_next_provider_call_finishes`),
     and the final page must be atomic with the terminal status (guarded by
     `tests/integration/test_pipeline_atomicity.py`). **Do not collapse the whole batch into one
     transaction** to simplify things — that breaks multi-page durability.

2. **`AppError` gained an internal-only `detail` field.** `message` is user/API-facing and must
   stay generic; `detail` carries the root cause and flows into evidence records via
   `format_error_detail` and into logs. Documented in `docs/error_handling.md` §"Message vs
   detail split". **When adding error paths: never put exception text into `message`.**

3. **8 `ui.notify` error sites replaced with `show_error`** in `home_page.py` / `people_page.py`.

4. **New AST guard** `test_ui_boundaries.py::test_no_page_hand_rolls_error_notifications`.

---

## 5. Phase 2 — Enforcement hardening

### P2-1 · `ty` suppression policy, then make the hook blocking
**Report ref:** LOW-05 · **Effort:** M

`.pre-commit-config.yaml` currently runs `ty` in **advisory** mode (a Python subprocess wrapper
that forces `sys.exit(0)`) because of the 10 known false positives. Net effect: a genuine new
type error prints alongside the known 10 and **blocks nothing**.

1. Add a targeted `# ty: ignore[<rule>]` at each of the 10 sites, each with a one-line comment
   explaining it's a SQLAlchemy descriptor false positive.
2. Confirm `uv run ty check` reports **0**.
3. Flip the pre-commit hook to blocking (drop the `sys.exit(0)` wrapper).
4. Document the policy in `docs/` — when a suppression is acceptable and what the comment must say.

**Acceptance:** `uv run ty check` → 0 diagnostics; introducing a deliberate type error fails
`git commit`; revert the deliberate error afterward.

### P2-2 · `ruff format` enforcement
**Report ref:** LOW-06 · **Effort:** S

~35 files have formatting drift. **Two separate commits, in this order:**
1. `uv run ruff format .` — formatting only, **no other changes in this commit**.
2. Add `ruff format --check` to `.pre-commit-config.yaml`.

Keeping these separate matters: a mixed commit makes the formatting noise unreviewable.

**Acceptance:** `uv run ruff format --check .` clean; full suite still 381+.

### P2-3 · Enable ruff ruleset `G` (flake8-logging-format)
**Report ref:** LOW-09 · **Effort:** S

f-strings in logging calls format eagerly regardless of level and break template grouping in
structured backends. Known instance: `workflows.py:193`. Enable `G` in `ruff.toml`, then convert
offenders to `%s` lazy args: `logger.error("Job %s failed.", job.id)`.

**Acceptance:** `uv run ruff check .` clean with `G` enabled.

---

## 6. Phase 3 — Reliability & concurrency

### P3-1 · Periodic stale-job recovery
**Report ref:** MED-01 · **Effort:** M

`requeue_stale_processing_jobs` has exactly one caller — `app.py:79`, in the lifespan startup
handler. There is no runtime re-check. The threshold reuses `worker_provider_timeout_seconds`
(**30.0s**, `config.py:116`).

The failure mode: a job orphaned <30s before a fast restart fails the staleness predicate at the
only moment recovery runs, so it stays `PROCESSING` forever (the worker only claims `QUEUED`).
Restarts are exactly when orphans are created, so the recovery window is systematically
misaligned with the failure it exists to handle.

1. Add `worker_stale_job_seconds` to `Settings` (don't keep overloading the provider timeout —
   they need independent tuning). Update `.env.example` in the same change (required by
   `services.instructions.md`).
2. Run the sweep periodically in the worker loop, **in addition to** the startup call.
3. Test: a job left `PROCESSING` past the threshold is requeued **without** a restart.

### P3-2 · Gate retries on error category
**Report ref:** MED-02 · **Effort:** S

`workflows.py:184-194` gates only on `job.retry_count < settings.worker_max_retries`. It never
consults `error_category` or `AppError.retriable`, so `validation` / `not_found` / `conflict`
failures would retry to exhaustion, burning provider quota on calls that cannot succeed.

**Latent today** because `worker_max_retries` defaults to `0` — which is exactly why this must be
fixed *before* anyone raises that value. Add backoff too; retries currently requeue immediately.

**Acceptance:** a `validation`-category failure is not requeued even with `worker_max_retries=1`.

### P3-3 · Shutdown budget derived from provider timeout
**Report ref:** MED-04 · **Effort:** S

`worker.py:146` waits `2.0s` for the worker task, but an in-flight provider call may run 30s and
the stop event is only checked *between* jobs. Derive the budget from
`worker_provider_timeout_seconds` plus a small grace. Document the relationship to the container
termination grace period in `docs/production-runbook.md`.

### P3-4 · Handle `IntegrityError` on the attempt-number flush
**Report ref:** MED-03 · **Effort:** M

`sources.py:540-546` computes `MAX(attempt_number) + 1`; `uq_execution_attempt_number` enforces
uniqueness. The sibling `JobSource` insert catches `IntegrityError` at `sources.py:531-534`, but
the attempt `flush()` at `sources.py:587` does **not** — a race loses an evidence row.

Not reachable today (single worker, sequential sources). **It becomes reachable the moment a
second worker replica is deployed** — treat this as a hard precondition for horizontal scaling
and note that in `docs/production-runbook.md`.

Mirror the `JobSource` handling: catch, recompute, retry bounded, raise a domain error on
exhaustion.

### P3-5 · Move blocking work off the event loop
**Report ref:** LOW-01, LOW-02, LOW-03 · **Effort:** S

- `store.py:401` — `hashlib.sha256(file_bytes)` is CPU-bound on the loop. Wrap in `asyncio.to_thread`.
- `homepage_store.py:25,32` — sync file I/O called from async page handlers. Same fix.
- `app.py:62` — `poll_interval_seconds=1.0` hardcoded. Move to `Settings`; update `.env.example`.

Every other I/O path already uses `to_thread` (`media_storage.py:43`, `normalization.py:117`,
`photos.py:176`, `sources.py:740,753`) — follow those.

---

## 7. Phase 4 — Consolidation

### P4-1 · Single transaction entry point
**Report ref:** MED-05 · **Effort:** M

`workflows.py` opens transactions via `services.jobs._session_scope()` and
`services.sources._session_scope()` — **private members of two different services**. This is the
mechanism that made HIGH-01 easy to introduce: nothing in the design signals that two scopes are
being opened for one logical unit of work.

Introduce `unit_of_work(services, session)` (proposed signature in review §7), migrate
`workflows.py` onto it, then add an AST guard to `test_service_boundaries.py` forbidding
`_session_scope` access outside its owning module. Note the existing test checks *imports*, not
attribute access, so it can't currently see this.

**Do not attempt this before Phase 1's tests are green in your working tree** — it touches the
same functions.

### P4-2 · Extract shared helpers
**Report ref:** §7 · **Effort:** M

`run_blocking` and `insert_with_sequence_retry`, per the review's proposed signatures. Do this
*after* P3-4 and P3-5, so the call sites exist.

### P4-3 · Prune low-signal tests
**Report ref:** LOW-10, LOW-11 · **Effort:** S

Full table in review §6 "Prune/strengthen backlog". Highlights:
- `test_traceability.py:54-57` — asserts properties of dict literals in the same file.
- `test_pipeline_flow.py:135-140,446-452` — `assert processed is True` is unfalsifiable
  (`read_job` raises rather than returning `None`).
- `test_orphan_sweep.py:119` — `>= 200` snapshot threshold tolerates ±40 drift.
- `test_workflows_reliability.py:157-196` — real `time.sleep(0.40)` with only 10% slack; flaky
  under CI load.

---

## 8. Phase 5 — Governance

- **P5-1** — Extend `test_orphan_sweep.py` beyond module-level definitions to public methods
  (LOW-08); seed `KNOWN_ORPHANS` with current results to keep it non-breaking.
- **P5-2** — Resolve the 4 "uncertain" orphans (LOW-07). Note `summarize_error` should now be
  reachable — consider using it, or delete it.
- **P5-3** — Log incomplete request manifests instead of silently returning `None`
  (`openrouter.py:347`, LOW-04).
- **P5-4** — Consider inverting the UI boundary check from a forbidden-list to an allowlist; a
  future persistence helper under a new name currently escapes it.

---

## 9. Ground rules

1. **Red first.** For any behavioral fix, write the test, *run it, observe it fail*, then fix.
   That's how Phase 1 caught that its own first HIGH-03 attempt was wrong.
2. **One phase per commit series.** Don't mix P2-2's formatting sweep with logic changes.
3. **Docs in the same change.** Required by `documentation-sync.instructions.md` whenever
   contracts, behavior, or `Settings` change. `Settings` changes additionally require
   `.env.example` updates in the same commit.
4. **Don't widen the NiceGUI pin.** `nicegui==3.13.0` is a deliberate release-stability decision
   recorded in `docs/production-runbook.md`. It is explicitly **not** a defect.
5. **`docs/reviews/**` is not canonical.** It's a dated artifact; don't treat it as a contract
   the way `docs/schema.md` or the instruction files are.
6. **Ask before scope-expanding.** If a fix seems to require restructuring beyond its task,
   stop and confirm — that's the signal a Phase boundary is being crossed.

## 10. Suggested first command

```powershell
cd C:\Github\transcription
git log --oneline -3
uv run ruff check . ; uv run pytest -q -m "not external" ; uv run ty check
```

Confirm the baseline (clean ruff, 381+ passing, exactly 10 `ty` diagnostics) before changing
anything. If that doesn't reproduce, stop and report rather than proceeding.