fix(checks): catch digit-level account and letter confusions

[AshSgDe29071999]

Summary

Mis-hearing checks covered the classic -teen/-ty amount pairs, but the expensive phone-line failures are digit and letter swaps on account numbers, card readbacks, and reference codes. _CONFUSABLE also was never consulted.

Changes

• Use _CONFUSABLE in check_misheard so a matching pair is misheard_number / high
• Add oh/zero, a/eight, to/too/two, double-N vs -ty N, and B/D/E/P/T/V/Z pairs with acoustic comments
• Fixture fixtures/misheard_account.json + test
• Existing good_call.json still has zero findings

Test plan

• python3 -m pytest tests/test_voiceeval.py -q (20 passed)

Fixes #4

[royalpinto007]

Good fix, and the tests pass. Verified the dropped-sign bug was real.

One change needed: _confusable_hit checks whether one side appears anywhere in the transcript and the other anywhere in the truth, with no position check. Running your branch:

FLAGGED  "I need to transfer two hundred pounds"   <- transcript == truth
FLAGGED  "I have a card ending eight"              <- transcript == truth

to and a are everywhere in English, so this fires on clean calls. good_call.json only passes because it happens to contain no "two" or "eight".

Fix: compare position, not presence, using difflib.SequenceMatcher opcodes and only flagging tokens that actually changed. Worth adding a fixture too: a clean call containing both halves of a pair, asserted to give zero findings.

[royalpinto007]

The confusable-pair check can produce high-severity false positives when both transcripts contain the same number but differ elsewhere. Please ensure it detects an actual token substitution and add a regression test for this case.