Fix/episodic capture timestamp and dedup guard

[joerybka]

Summary

• Fix _find_transcript_ts's content-hash fallback in migrate --rederive-timestamps so it no longer matches an internal transcript event (e.g. queue-operation) ahead of the real user/assistant turn when both carry identical text.
• Add a regression test reproducing a real collapsed-timestamp group where this caused the migration to silently re-derive the same already-collapsed timestamp on every run.

Motivation

iai-mcp doctor's collapsed-timestamp check can report groups that migrate --rederive-timestamps never clears, even though the migration reportsrecords_updated > 0 each run. Root cause: the migration's content-hash matcher scans transcript lines top-to-bottom and returns on the first hash match, with no filter on line type. When an internal event (e.g. queue-operation) carries the same duplicated text as the real turn and appears earlier in the transcript file, the matcher locks onto it and returns its timestamp instead — which is the same collapsed value the record already has. The write is a no-op, but the function still counts it as updated.

The original capture path (capture_deferred_captures in capture.py) already filters transcript lines to type/role in {"user", "assistant"} before considering them; the migration's independent matcher never had that guard.

Changes

• src/iai_mcp/migrate/_timestamps.py: filter candidate transcript lines to user/assistant type/role before content-hash comparison, mirroring the existing guard in capture.py.
• tests/test_migrate.py: new test test_migrate_rederive_skips_internal_event_duplicate — builds a collapsed group of 3 duplicate episodic records with no source_uuid (forcing content-hash fallback) against a transcript containing an earlier queue-operation line and a later real user line with identical text, and asserts the migration resolves to the real turn's timestamp, not the collapsed one.

Test plan

• pytest tests/test_migrate.py — 25 passed
• Confirmed the new test fails against the pre-fix matcher and passes with the fix
• ruff check / ruff format --check on changed files show only pre-existing findings already present on main, unrelated to this diff

No bench re-run included — this is a data-repair migration correctness fix, not a retrieval/capture/consolidation performance change.

Note

This doesn't fully resolve doctor's collapsed-timestamp warning for groups that are true duplicate records (same content+provenance) rather than distinct turns sharing a bad timestamp — those still need migrate --dedupe-episodic to collapse the duplicates themselves. This PR only fixes the matcher so rederive stops silently no-op'ing on genuinely distinct turns that happen to share text with an internal event.