Fix selective retranslation for untracked glossary-sensitive chunks

[cursor]

Bug and impact

Existing output_chunk*.md files without run_state.json records were adopted as record-only before checking whether the current glossary affected the chunk. A stale translation could be recorded as if it already used the current glossary, then skipped forever as unchanged. Alias-only glossary edits for already-selected terms could also be missed because aliases were not part of the per-term hash.

Root cause

run_state.plan() short-circuited untracked existing outputs to record before glossary term selection/hash checks, and glossary.term_hash() only hashed source, target, and category.

Fix

• Re-translate untracked existing outputs when the current glossary selects terms for that chunk, while preserving record-only adoption for glossary-free chunks.
• Include sorted aliases in term_hash() when aliases are present, preserving no-alias hash compatibility.
• Update SKILL/README docs for the corrected default behavior.
• Add regression coverage for untracked glossary-sensitive outputs and alias-only changes.

Validation

• Reproduced the pre-fix CLI scenario where stale chunk0001 was recorded and then marked unchanged.
• Post-fix CLI scenario keeps glossary-sensitive chunk0001 in translation_chunk_ids through record-all, while record-only adopting glossary-free chunk0002.
• python3 -m unittest tests.test_run_state tests.test_glossary -v (66 tests OK)
• python3 -m unittest discover -s tests -p 'test_*.py' -v (207 tests OK, 3 skipped)
• python3 -m compileall scripts tests