feat: detect confirmed wrong values in voice calls

[Amidwestnoob]

What & why

Closes #20. A confirmation turn can echo a misheard amount and still satisfy the existing missing-confirmation check. Add a sibling confirmed_wrong_value check to flag contradictory ground-truth or bounded action evidence.

Changes

• Recognize relevant confirmation phrases, including the issue's “Confirming fifty?” example, and normalize supported isolated English words and unsigned digit/decimal values.
• Bound associations to the current caller exchange to avoid unrelated later requests and duplicate findings. Handle mixed correct/incorrect numbers and unambiguous action amounts.
• Add a wrong-confirmation fixture, a dedicated regression suite and README coverage/limitation notes. The original test suite remains unchanged.

Validation

Independent Astra High review and repairs; Python 3.12.13:

python -m pytest -q
ruff check voiceeval tests
ruff format --check voiceeval/checks.py tests/test_confirmed_wrong_value.py

65 tests passed. Lint passed with Ruff 0.15.12. Changed-file formatting passed. Separate CLI assertions verified clean versus flagged fixture exit codes and output. Additional regression cases failed on the submitted version and passed after repair.

Ruff 0.16.7 reports the same five inherited diagnostics on base and final. Whole-tree formatting reports the same three unchanged files on both. Those baseline failures were preserved rather than suppressed or included in an unrelated cleanup. Python 3.11 and live STT/provider/call testing were not run locally.

Limitations

This is a bounded English heuristic, not a general spoken-number or semantic-field parser. Compound/signed/grouped forms are conservatively skipped where detected. Locale packs, word “zero”, numeric repetition/order and arbitrary confirmation language are not supported. A ground-truth contradiction flags the confirmation itself; it does not assert caller acceptance or an executed transaction. The README documents these boundaries.

Checklist

• Focused change with a clear description
• Ran relevant format, lint and tests; baseline limitations noted above
• Updated README with check behavior and numerical coverage limits
• Linked the related issue

AI-assisted contribution, independently reviewed, repaired and tested by a separate Astra High subagent.