skill shorten loses its oracle work when --max-usd stops it #630

skill shorten is wrapped by stops_at_cap (dokimi/dokimi/commands/skill.py:166), which turns SpendCapReached into SystemExit("...; nothing was written") (commands/novelty_run.py:37). analyze, atoms and probe keep a partial result at the cap; shorten does not.

Repro (mocked client, offline): save an atoms file for 30 spans whose statements only partly overlap their spans, then run skill shorten --artifact rows.md --atoms atoms.json --model o --out short.md --max-usd 0.05 with a client whose cap is already spent. Output: EXIT: spend cap reached: $0.05 of $0.05 after 0 calls (next call: o). Raise --max-usd, or measure less (a cheaper --tier, a lower --max-atoms); nothing was written, and the directory holds atoms.json and rows.md only.

What shorten pays for is retention-oracle verdicts. Options: keep the verdicts already obtained (--cache-dir already stores them; say so in the cap message), or score the candidates with what was adjudicated and mark the table partial. cmd_novelty_shorten is dokimi/dokimi/commands/novelty.py:451.

Found while fixing #624. The fix's PR is #633.