feat(auto): persist source-tagged auto answer log

[shaun0927]

Summary

Closes the last #639 surface gap: "Interview progress shows round count and source-tagged auto answer summaries."

Records each auto-answered interview round on AutoPipelineState.auto_answer_log as a small dict (round, AutoAnswerSource tag, truncated question/answer). ooo auto --status renders the most recent five entries so users can see why the auto pipeline made each answer choice without scraping `~/.ouroboros/data/auto_*.json`.

Changes

• src/ouroboros/auto/state.py
  • New field auto_answer_log: list[dict[str, Any]] = field(default_factory=list).
  • from_dict backfills missing field to [] for legacy persisted sessions.
  • _validate_loaded extends the existing list-of-dicts check (previously only findings) to cover auto_answer_log.
• src/ouroboros/auto/interview_driver.py
  • New module-level helper _record_auto_answer(state, *, round_number, source, question, answer) appends to the log right after each successful answer is selected and capped at the last 25 entries.
  • _truncate(text, limit) collapses whitespace and trims long Q/A text to 200 chars.
  • Constants _AUTO_ANSWER_LOG_LIMIT = 25 and _AUTO_ANSWER_LOG_TEXT_LIMIT = 200.
• src/ouroboros/cli/commands/auto.py
  • _print_status renders `Recent auto answers (last N):` followed by `round X \[source] Q: ...` / `A: ...` lines, when the log is non-empty. The leading [ is escaped so Rich treats [source] as literal text.
• tests/unit/auto/test_answer_source_log.py (new):
  • Default empty list, round-trip persistence, legacy hydration, non-list rejection.
  • _record_auto_answer truncation and ring-buffer eviction.
  • _truncate whitespace collapsing.
  • End-to-end CLI `--status` rendering.
  • CLI omits the `Recent auto answers` block when the log is empty.

Pre-flight: PR collision check

The other in-flight #675 work (#682 "add selected-driver answer metadata" and #683 "classify selected-driver answer risk") touches src/ouroboros/auto/answerer.py, src/ouroboros/auto/driver_answerer.py, src/ouroboros/auto/__init__.py, and tests/unit/auto/test_driver_answerer.py. This PR touches none of those files. PR-E intentionally records only the existing AutoAnswer.source enum tag rather than the richer AutoAnswerMetadata introduced by #682, so the two efforts are orthogonal and can land in either order. A future slice can extend the log entry shape with AutoAnswerMetadata fields once #682 / #683 land.

Validation

• UV_CACHE_DIR=/tmp/uv-cache uv run ruff check src/ouroboros/auto/state.py src/ouroboros/auto/interview_driver.py src/ouroboros/cli/commands/auto.py tests/unit/auto/test_answer_source_log.py → passes.
• UV_CACHE_DIR=/tmp/uv-cache uv run pytest tests/unit/auto -q → 208 passed.

Risks / Notes

• Field is additive; legacy sessions hydrate cleanly with an empty log.
• The bounded ring-buffer keeps the persisted JSON compact even on long-running sessions (worst case: 25 × ~500 bytes ≈ 13 KB).
• The CLI rendering is intentionally conservative (last 5 entries) to avoid flooding terminal output. A future PR can add --show-answer-log for the full log if needed.

Refs #639

[ouroboros-agent]

Review — ouroboros-agent[bot]

Verdict: REQUEST_CHANGES

Reviewing commit 642e158 for PR #720

Review record: 12ae5c42-2226-4ccd-8b87-e7c6c94d1acf

Blocking Findings

| # | File:Line | Severity | Finding |
|### Recovery Notes
First recoverable review artifact generated from codex analysis log.

---|-----------|----------|---------|
| 1 | src/ouroboros/cli/commands/auto.py:288 | BLOCKING | The new status output prints question and answer through console.print() without escaping Rich markup. These fields come from interview/backend text and can legitimately contain [ / ]; when that happens Rich will interpret them as markup, which can corrupt the displayed text or raise a markup parse error and make ooo auto --status fail for persisted sessions that otherwise load correctly. |

Non-blocking Suggestions

None.

Design Notes

The persistence change is small and backwards-compatible, and bounding the saved log is the right shape for a resume-oriented state file. The main issue is that the new CLI surface treats stored interview text as trusted Rich markup instead of plain content.

---

Reviewed by ouroboros-agent[bot] via Codex deep analysis

[Q00]

Reviewed. Persisting the source-tagged answer log touches state, CLI, and tests, so it carries regression risk. CI is currently failing, so I cannot approve it yet.

[ouroboros-agent]

Review — ouroboros-agent[bot]

Verdict: APPROVE

Reviewing commit 46276ba for PR #720

Review record: 734ce955-a417-4311-bf1c-962edf74aa97

Blocking Findings

No in-scope blocking findings remained after policy filtering.

Non-blocking Suggestions

None.

Design Notes

The change is narrowly scoped and consistent with the existing durable-session design: answer provenance is recorded at the point of state mutation, persisted with backward-compatible hydration, and rendered defensively in the CLI to avoid Rich markup breakage from backend-provided text.

Recovery Notes

First recoverable review artifact generated from codex analysis log.

---

Reviewed by ouroboros-agent[bot] via Codex deep analysis

[shaun0927]

@Q00 re-review ping.

This PR is the final scope-cleanup follow-up for #639 — it lands the source-tagged auto answer log that the issue's acceptance criteria ("Interview progress shows round count and source-tagged auto answer summaries") explicitly call for, but in a way that is intentionally orthogonal to the in-flight #682 / #683 `AutoAnswerMetadata` work so the two efforts can land in either order without conflict.

What changed since the first round of review

The original blocking finding was that the new CLI status surface printed the persisted question / answer text directly through Rich's markup parser, so any backend response that legitimately contained `[` / `]` could either be silently rewritten as a non-existent style or raise a markup parse error and break `ooo auto --resume --status`. The fix wraps the source / question / answer fields with `rich.markup.escape()` before printing, and a new regression test feeds a mixed-bracket payload (`Use [bold]uv[/] toolchain` + `Use uv [from existing setup]`) end-to-end through the status surface and asserts both literal bracketed segments survive verbatim.

The latest deep review (commit `46276ba`) accepted the change with no remaining blocking or non-blocking findings:

Verdict: APPROVE

The change is narrowly scoped and consistent with the existing durable-session design: answer provenance is recorded at the point of state mutation, persisted with backward-compatible hydration, and rendered defensively in the CLI to avoid Rich markup breakage from backend-provided text.

Why this is safe to merge

• Additive on AutoPipelineState: `auto_answer_log` defaults to `[]`, hydrates legacy persisted sessions to `[]`, and is bounded to the last 25 entries (≈ 13 KB worst case), so existing sessions and consumers are unaffected.
• Single producer surface: the log is appended only inside `AutoInterviewDriver` at the existing post-answer save site, with no churn to `auto/answerer.py`, `auto/driver_answerer.py`, `auto/init.py`, or `tests/unit/auto/test_driver_answerer.py` — the exact files feat(auto): add selected-driver answer metadata #682 / feat(auto): classify selected-driver answer risk #683 are touching — so PR-682/feat(auto): classify selected-driver answer risk #683 can land in either order.
• CI green across Ruff lint + format, MyPy, and pytest on Python 3.12 / 3.13 / 3.14, plus Bridge TypeScript.
• CLI is robust to arbitrary backend text, so a malformed interview reply cannot break the operator-facing status surface.

Refs #639. Ready for human review.

[Q00]

Reviewed across OS/UserLevel/Program boundaries, auto scope, and UX complexity. Approving: the answer-source log is additive diagnostic state, bounded in status output, and helps users audit auto-filled interview answers without changing the auto decision model or adding a new command path.