Fix run_state adoption of glossary-sensitive stale outputs

[cursor]

Bug and impact

Untracked output_chunk*.md files were accepted as record-only even when their source chunk selected terms from the current glossary. A stale translated chunk could be recorded under the new glossary and then become permanently unchanged, silently preserving wrong terminology in final book outputs. Alias-only edits to an already-selected term were also ignored because aliases were omitted from the per-term hash.

Root cause

run_state.plan() treated every existing output without a run_state.json record as safe to adopt unless --retranslate-untracked was passed. glossary.term_hash() hashed only source/target/category even though aliases are injected into subagent prompts.

Fix

• Require untracked chunks that select current glossary terms to be translated before recording state.
• Keep glossary-insensitive untracked outputs eligible for record-only adoption.
• Include sorted aliases in term hashes when aliases are present.
• Update SKILL/README docs and add regression tests for both stale adoption and alias-only glossary changes.

Validation

• Pre-fix repro showed chunk0001 record-only, then permanently unchanged after record-all and alias edits.
• Post-fix repro shows chunk0001 in translation_chunk_ids, chunk0002 record-only, and alias edits retrigger chunk0001.
• python3 -m unittest tests.test_run_state tests.test_glossary -v (68 tests OK)
• python3 -m unittest discover -s tests -p 'test_*.py' -v (209 tests OK, 3 skipped)
• python3 -m compileall scripts tests
• Standard Alice mock full-pipeline smoke: 38 chunks merged; HTML/DOCX/PDF generated successfully; EPUB failed only with the documented Ubuntu Calibre 7.6.0 bytes/str issue.