[kiranjd]

We hit these bugs while dictating with Speakmac, which uses this crate for on-device number formatting:

Spoken	Before	After
she is forty-six years old	she is forty-six years old	she is 46 years old
the thirty-second floor	the thirty-second floor	the 32nd floor
add one and three quarters cups of flour	add 13/4 cups of flour	add 1 3/4 cups of flour
wait two and a half hours	wait 2 half hours	wait 2 and a half hours
the score was forty two to thirty seven	the score was 36:18	the score was 42 to 37
I have twenty one apples and thirty two oranges	I have 21 apples 32 oranges	I have 21 apples and 32 oranges
the total was three thousand nine hundred nine	the total was 3900 9	the total was 3909
the meeting is at three pm.	the meeting is at 03:00 p.m**..**	the meeting is at 03:00 p.m.
(fr) le cent onzième jour	le 100/11 jour	le 111ᵉ jour
(fr) quarante pour cent	40 pour 100	40 %
(de) vierzig prozent	40 prozent	40 %

Two existing tests asserted the old output, and their comments already flagged it as a known issue: test_sentence_mixed_punctuation_and_numbers and test_sentence_itn_multiple_numbers. Both are updated.

How

• en: hyphens count as word breaks. Cardinal spans can't start or end on "and"/"a". Mixed fractions split on "and" ("a half" stays as words). "X to Y" times need a valid hour and minute.
• Sentence loop: removed the 4-token cardinal limit, since max_span_tokens already bounds the search. A single period right after a tagger output that already ends in "." (such as "p.m.") is dropped; ellipses are kept.
• fr/de: a singular denominator can't take a numerator above one. Added pour cent and prozent → %.

Testing

• cargo test: 1,172 passed, 0 failed. cargo test --features ffi: 1,180 passed. cargo fmt --all --check is clean, and cargo clippy --all-targets reports no new warnings.
• New unit tests cover each tagger change. New sentence-level tests in en_tests.rs and multilang_itn_tests.rs cover every row in the table above.
• examples/parity against NeMo at the pinned ref 1f12635 (tests/parity_baseline.tsv updated):
  • fr ITN ordinal goes from 22/23 to 23/23, which brings fr ITN overall from 306/307 to 307/307.
  • Every other language and class is unchanged, including all TN classes.

Scope

• No public API, FFI or wasm changes. All of this is default-path behaviour.
• Happy to split this into per-language PRs if that's easier to review.

🤖 Generated with Claude Code

[Alex-Wengg]

Thanks, this reproduces cleanly here and the parity run is clean. One blocker: the "oh" branch in parse_span turns bare times into digits in sentence mode ("meet me at nine oh five" → "905", main gives "09:05"), while single-expression mode still gives 09:05. Could you drop that branch and its test from this PR (room numbers need preceding-token context, happy to take that as a follow-up) and refresh tests/parity_baseline.tsv for the fr ordinal 22→23? The rest looks good to merge.

[kiranjd]

Thanks for catching that. Pushed 7bfaee7:

• Dropped the "oh" branch and its test. Added test_sentence_oh_time_stays_a_time
• Refreshed tests/parity_baseline.tsv
• cargo test: 1,172 passed