feat(api): enrich refresh_mental_model result_metadata with semantic outcome

[hb-cam]

Fixes #2605.

What

refresh_mental_model operations completed with result_metadata carrying only the submit-time {mental_model_id, name} stub (set in submit_async_refresh_mental_model before the op runs, never enriched). This PR mirrors the retain enrichment pattern (operation_metadata.py + completion-time metadata write) as suggested in the issue thread: the worker handler now merges the semantic outcome into result_metadata at completion.

New fields on completed refresh ops:

• content_len — length of the final stored content
• populated_content — true only for real synthesis: the reflect agent's "No answer provided." fallback and the "Generating content..." placeholder complete wire-successful but read as false (a bare length check would report the 19-char stub as populated)
• based_on_counts — per-fact-type grounding counts from the reflect response

How

• RefreshMentalModelOutcomeMetadata dataclass alongside RetainOutcomeMetadata in operation_metadata.py
• _write_refresh_outcome_metadata in the engine, mirroring _write_retain_outcome_metadata: jsonb merge (COALESCE(result_metadata, '{}'::jsonb) || …) so the submit-time keys consumers join on are preserved, not replaced; best-effort with a loud warning on failure
• Called from _handle_refresh_mental_model after a successful refresh, before the worker marks the operation completed (same ordering as retain)
• The reflect agent's fallback literal is promoted to a NO_ANSWER_TEXT constant so the populated judgment compares against the constant rather than a copied string

No endpoint or request model changes — result_metadata is already a free-form dict in the operation status response, so no OpenAPI/client regeneration is needed.

Testing

• New tests/test_refresh_outcome_metadata.py (2 tests, written first and watched fail): a completed refresh exposes all three fields with the submit-time keys intact; the no-answer stub completes but reads populated_content: false
• Tests exercise the real path (submit → task backend → handler → metadata write → get_operation_status) against the embedded pg0 database with only the LLM-bound refresh_mental_model stubbed
• Neighboring suites pass: test_mental_model_scheduled_refresh.py, test_mental_model_delta.py, test_mental_model_hooks.py (24 passed, 4 skipped), test_async_batch_retain.py (22 passed)
• ruff check, ruff format --check, ty check clean on changed files

🤖 Generated with Claude Code

[koriyoshi2041]

I checked this locally on macOS with Python 3.11.14 and the behavior looks good to me.

What I ran:

• git diff --check origin/main...HEAD
• env -u ALL_PROXY -u all_proxy -u HTTPS_PROXY -u https_proxy -u HTTP_PROXY -u http_proxy uv run pytest tests/test_refresh_outcome_metadata.py tests/test_mental_model_scheduled_refresh.py -q -n0

Result: 8 passed, 2 warnings in 21.83s. The first run inherited my local SOCKS proxy env and failed before exercising the patch because socksio is not installed; rerunning without proxy env cleared that and the focused refresh tests passed.

The extra no-answer stub coverage here is useful, so I am going to close my overlapping #2621 and leave this as the stronger path for #2605.