Fix stale glossary-sensitive outputs in selective retranslation

[cursor]

Bug and impact

Existing translated chunks without run_state.json records were treated as record_only by default, even when the current glossary selected terms for that chunk. The skill records record_only_chunk_ids before launching subagents, so a stale translation from an old glossary could be stamped as current and then skipped forever. Alias-only glossary edits also did not change per-term hashes, so already-selected chunks could miss prompt-affecting alias changes.

Root cause

run_state.plan() did not inspect selected glossary terms before adopting untracked outputs, and glossary.term_hash() only hashed source/target/category, omitting aliases that are injected into the prompt term table.

Fix

• Re-translate untracked outputs when the current glossary selects any terms for that chunk, while still record-only adopting glossary-free untracked outputs for resumability.
• Include sorted aliases in term_hash() when aliases are present, preserving hash compatibility for no-alias records.
• Update tests and docs/SKILL guidance for the refined behavior.

Validation

• Reproduced the pre-fix CLI bug: glossary-sensitive chunk0001 was record-only and became unchanged after record-all; alias-only change was unchanged.
• Post-fix CLI repro: chunk0001 stays in translation_chunk_ids, only glossary-free chunk0002 is record-only; alias-only change triggers translation.
• python3 -m unittest tests.test_run_state tests.test_glossary -v (68 tests)
• python3 -m unittest discover -s tests -p 'test_*.py' -v (209 tests, 3 skipped)
• python3 -m compileall scripts tests
• Standard Alice mock full pipeline: 38 chunks merged; HTML/DOCX/PDF generated. EPUB failed only with the documented Ubuntu Calibre 7.6.0 bytes/str issue.