Fix spelled decimals (#15) + add opt-in aviation pipeline (#14)

[Alex-Wengg]

Summary

Two related issues, two different fix shapes:

• Issue Failed on case "frequency one three five point six two five" #15 — spelled decimals (e.g. aviation radio frequencies): "one three five point six two five" returned 9.625 instead of 135.625. Default-path bug. Fixed in the default pipeline.
• Issue normalize_sentence failed on "United seven eighty eight" #14 — flight-number reading: normalize_sentence("United seven eighty eight") returned "United 95" (7+80+8) instead of "United 788". Genuinely ambiguous (e.g. "two thirty five" is 02:35 to a clock-reader, 235 to ATC, 65 to a math student). Fixed as an opt-in pipeline so callers in flight-number / call-sign context get aviation reading without forcing it on everyone.

Issue #15 — default-path fix

cardinal::words_to_number now dispatches between two readings:

1. Digit-by-digit — every token is a single-digit word (zero-nine, plus oh/o). Concatenate. This is the path decimal::parse_point_decimal reaches when reading the integer side of a spelled decimal.
2. Grammatical — original running-sum + scale-multiplication accumulator. Unchanged from upstream NeMo.

decimal::parse_point_decimal and telephone::parse were updated to use the digit-by-digit path. No public API change for #15.

Issue #14 — opt-in aviation pipeline

Why not auto-apply: I tried wiring aviation reading directly into words_to_number, and it broke date::parse_old_year for "twenty one forty two" (was 2043, became None because year_part started reading 142 > 99). Aviation context can't be detected purely from the tokens, so the only safe option is to let the caller decide.

New public API (additive)

Function	Purpose
cardinal::words_to_number_aviation(s)	Low-level: digit-prefix + grammatical-compound reader. "seven eighty eight" → Some(788). Falls back to grammatical when no digit prefix.
cardinal::parse_aviation(s)	String wrapper, mirrors cardinal::parse.
normalize_aviation(s)	Single-input dispatch. Aviation cardinal tried before time/date; falls through to standard normalize for non-number phrases.
normalize_sentence_aviation(s)	Sentence-mode dispatch. Aviation cardinal at priority 89 (above date=88, time=85; below measure=90, money=95).
normalize_sentence_aviation_with_max_span(s, n)	Configurable-span variant.

Internal change

parse_span now takes an aviation: bool parameter. Single tagger list, two short conditional blocks for the aviation injection. normalize_sentence_with_max_span passes false; the aviation variant passes true. No fork to maintain.

fn parse_span(span: &str, aviation: bool) -> Option<(String, u8)> {
    // ... money / measure ...
    if aviation && token_count <= 4 {
        if let Some(r) = cardinal::parse_aviation(span) { return Some((r, 89)); }
    }
    // ... date / time / electronic / decimal / ordinal ...
    if !aviation && token_count <= 4 {
        if let Some(r) = cardinal::parse(span) { return Some((r, 70)); }
    }
    None
}

Behaviour matrix

	normalize	normalize_aviation	normalize_sentence	normalize_sentence_aviation
"twenty one forty two"	2043	63	2043	63
"seven eighty eight"	788*	788	95	788
"two thirty five"	02:35	235	02:35	235
"United seven eighty eight"	788*	788	United 95	United 788
"flight two thirty five departs ..."	235-4*	235-4*	flight 02:35 ...	flight 235 ...
"one three five point six two five"	135.625	135.625	135.625	135.625
"five dollars"	$5	$5	$5	$5
"twenty first"	21st	21st	21st	21st
"two thousand seventeen"	2017	2017	2017	2017

* normalize single-input mode lets the telephone tagger eat the whole span; sentence mode does not.

Default normalize and normalize_sentence outputs are bit-for-bit unchanged from before this PR (apart from the #15 spelled-decimal fix).

Tests

Unit tests in src/itn/en/cardinal.rs:

• test_words_to_number_aviation_flight_number
• test_words_to_number_aviation_falls_back_to_grammatical
• test_words_to_number_aviation_scale_word_forces_grammatical
• test_words_to_number_no_aviation_reading — guards that generic words_to_number does not do aviation reading

Integration tests in tests/en_tests.rs:

• test_issue_14_default_dispatch_unchanged
• test_issue_14_normalize_aviation
• test_issue_14_normalize_sentence_aviation
• test_issue_15_* (already on branch)
• test_spelled_digit_cardinal* (already on branch)

Test plan

• cargo test — full suite passes (911 tests across 12 binaries)
• cargo fmt --check — clean
• Issue normalize_sentence failed on "United seven eighty eight" #14 reporter's exact input via opt-in: normalize_sentence_aviation("United seven eighty eight") → "United 788"
• Issue Failed on case "frequency one three five point six two five" #15 reporter's exact input: normalize("one three five point six two five") → "135.625"
• Regression: normalize_sentence("twenty one forty two") → "2043" (date), normalize("two thousand seventeen") → "2017"
• Money / measure / decimal / ordinal preserved in aviation pipeline ("five dollars" → "$5", "twenty first" → "21st", "five point two" → "5.2")

Closes #15.
Addresses #14 via new opt-in pipeline (default behaviour intentionally unchanged).

[devin-ai-integration]

Devin Review found 3 new potential issues.

View 15 additional findings in Devin Review.

[devin-ai-integration]

🟡 Uses unwrap() in library code, violating AGENTS.md style rule

The AGENTS.md specifies: "Error handling: prefer Option/Result; avoid unwrap()/expect() in library code". Three new call sites in src/itn/en/cardinal.rs use .unwrap() on single_digit_char(w) (lines 191, 227, 246). While these calls are guarded by a prior .all(|w| single_digit_char(w).is_some()) check, making them safe at runtime, they can be trivially replaced with .filter_map(|w| single_digit_char(w)) to satisfy the style guideline without any behavioral change.

Suggested change

	return words
	.iter()
	.map(|w| single_digit_char(w).unwrap())
	.collect::<String>()
	.parse()
	.ok();
	return words
	.iter()
	.filter_map(|w| single_digit_char(w))
	.collect::<String>()
	.parse()
	.ok();

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🟡 Uses unwrap() in library code (words_to_number_aviation digit-by-digit path)

Same AGENTS.md rule violation as BUG-0001, in the words_to_number_aviation function's digit-by-digit reading path at src/itn/en/cardinal.rs:227. Guarded by .all() but can be replaced with .filter_map().

Suggested change

	return words
	.iter()
	.map(|w| single_digit_char(w).unwrap())
	.collect::<String>()
	.parse()
	.ok();
	return words
	.iter()
	.filter_map(|w| single_digit_char(w))
	.collect::<String>()
	.parse()
	.ok();

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🟡 Uses unwrap() in library code (aviation prefix concatenation path)

Same AGENTS.md rule violation, in the aviation prefix concatenation logic at src/itn/en/cardinal.rs:246. The .map(|w| single_digit_char(w).unwrap()) can be replaced with .filter_map(|w| single_digit_char(w)).

Suggested change

	let prefix: String = words[..prefix_len]
	.iter()
	.map(|w| single_digit_char(w).unwrap())
	.collect();
	let prefix: String = words[..prefix_len]
	.iter()
	.filter_map(|w| single_digit_char(w))
	.collect();

---

Was this helpful? React with 👍 or 👎 to provide feedback.