fix(asr): Preserve names, possessives and surrounding words

[achembarpu]

Why is this change needed?

Vocabulary rescoring can duplicate already-correct names, remove possessive endings, or consume the word following a hyphenated full name.

Examples:

• mister Quilter becomes Quilter Quilter.
• mister Quilter's manner loses its possessive.
• Correcting Chandra Shankar-Solansky said deletes said.

What changes?

• Protect recognized vocabulary phrases and preserve singular possessives in scoring and replacement text.
• Bound acoustic search margins at separately timed occurrences of the same name.
• Prevent separate candidates from borrowing the same acoustic occurrence outside their source spans.
• Suppress positive bias for weak matches that would create adjacent duplicates, while retaining normal bias for close spellings and exact aliases.
• Allow shorter source spans when their hyphen-separated component count exactly matches a canonical phrase or registered alias, preserving the existing span ceiling and multi-word similarity floor.

Legitimate corrections remain available, including longer full-name corrections that preserve recognized prefixes. Coarse overlapping timestamps do not establish acoustic ownership. The production logic contains no special-case names or verbs.

Add licensed real-speech fixtures and regression coverage for name preservation, possessives, repetitions, aliases and hyphenated name boundaries.

Validation

• 176 XCTest and 10 Swift Testing checks pass in the final affected selection.
• ASR-only Release build and touched-file formatting pass.
• Real-audio comparisons cover seven recordings across four ASR models under app and SDK-default policies.
• Both short Quilter recordings remain verbatim under the app policy for all four models; legitimate Solanki corrections remain available.
• The name-boundary correction leaves 54 of 56 outputs unchanged and restores exactly said in the two affected v2 outputs.
• Actual app consumer validation reproduces the missing word before the boundary fix and preserves the complete expected transcript afterward. Existing v3/shared-110M preservation checks also pass.

The broader streaming-rescue test class still reproduces the existing upstream Hey → Codex false positive and was excluded from the passing selection. Other acoustic-rescue false positives remain outside this fix. No performance claim.

Dependencies

None.