fix: remine changed conversation transcripts

[AlekseiSeleznev]

Summary

• persist source_mtime for conversation transcript drawers and registry sentinels
• make convo_miner re-mine a transcript when the source file changed
• add regression coverage for a transcript that grows between mine runs

Problem

convo_miner currently skips a transcript after its first successful ingest unless the normalization schema version changes. That works for one-shot exports, but it drops new turns when the same transcript file continues to grow across later mine runs.

Reproduction

1. Mine a conversation directory.
2. Append another exchange to the same transcript file.
3. Run mempalace mine <dir> --mode convos again.
4. The second run skips the file and the appended exchange never gets filed.

Fix

Store source_mtime in conversation transcript metadata and reuse the existing check_mtime=True freshness gate from project mining. If the transcript changed, convo_miner purges and rebuilds drawers for that source file; unchanged transcripts are still skipped.

Validation

• ./.venv/bin/pytest tests/test_convo_miner.py tests/test_hooks_cli.py tests/test_save_hook_mines.py -q
• ./.venv/bin/ruff check mempalace/convo_miner.py mempalace/palace.py tests/test_convo_miner.py
• ./.venv/bin/ruff format --check mempalace/convo_miner.py mempalace/palace.py tests/test_convo_miner.py