Fix stale glossary adoption for untracked outputs

[cursor]

Bug and impact

Untracked existing output chunks could be recorded into run_state.json even when the current glossary would inject terms into that chunk. If a user edited glossary.json before the first run-state record existed, stale translations could be treated as current and merged into final artifacts without retranslation.

Root cause

run_state.plan() marked all valid untracked outputs as record by default unless --retranslate-untracked was passed. Glossary term selection/hash checks only ran after a chunk already had a run-state record.

Fix

• Require translation for untracked existing outputs that have selected glossary terms under the current glossary.
• Continue allowing record-only adoption for untracked chunks with no selected glossary terms.
• Update the skill docs and both READMEs to describe the safer adoption behavior.
• Add regression coverage for glossary edits before first run-state recording.

Validation

• python3 -m unittest tests.test_run_state -v
• python3 -m unittest discover -s tests -p 'test_*.py' -v (206 tests, 3 skipped)
• python3 -m compileall scripts tests
• CLI reproduction: constructed an old-output/no-run_state temp dir with current glossary target Tai -> 泰; scripts/run_state.py plan now returns translation_chunk_ids: ["chunk0001"] for the glossary-sensitive chunk and only record-only adopts the chunk with no glossary hit.