feat: autocomplete speaker name selection

[r3dbars]

Summary

• enables NSComboBox completion in the speaker review sheet
• extracts speaker label sorting/completion/selection into a tested policy helper
• preserves duplicate-name disambiguation labels while allowing unique prefix autocomplete

Test Plan

• bash run-tests.sh
• bash build.sh

User feedback: typing an existing speaker name should auto-complete/select it instead of requiring a dropdown click.

[r3dbars]

PR Review

Verdict: APPROVE

Summary: Clean extraction of speaker name selection logic from SpeakerRowView into a testable SpeakerNameSelectionPolicy enum, plus enabling NSComboBox inline autocomplete. The policy is stateless (caseless enum with static functions), well-tested, and the UI integration is minimal — one flag flip (completes = true) and delegate method wiring. Net -81 lines from the view, +129 in a focused policy file with +60 in tests.

Blockers: None.

Should fix:

• Subtle behavior change in duplicate detection: The old makeIdentityLabels grouped duplicates by $0.displayName.lowercased(). The new version uses normalizedSearchText(displayName($0)), which adds .diacriticInsensitive folding. This means "José García" and "Jose Garcia" would now be treated as duplicates and get disambiguation labels (with call count + UUID prefix), where previously they'd be treated as distinct. This is arguably better behavior, but it's a silent change worth being aware of — confirm this is intentional.

• "You" sentinel not in lookup: The "You" label is prepended to knownPeopleLabels (line 445) but never added to knownPeopleByLabel. If a user manually types "You" (without the batch "Keep as You" toggle), knownPeopleOption(matching: "You") returns nil, so buildUpdate() emits a .named action with newName: "You" rather than a .collapsedToMe. Verify the downstream SpeakerNamingCoordinator handles a .named update with newName: "You" gracefully — if it does, this is fine; if it expects .collapsedToMe for that case, it could create an orphan profile.

Nits:

• Every call site and test case repeats the same id: { $0.id }, displayName: { $0.displayName }, callCount: { $0.callCount } closure triple. A test helper or a lightweight protocol conformance could cut the boilerplate, but the generic approach is reasonable for keeping the policy type-agnostic.
• normalizedSearchText calls .folding(options: .caseInsensitive) then .lowercased(). The .lowercased() is mostly redundant after case-insensitive folding, though it serves as a safe normalization guarantee across locales — fine to keep.
• visibleKnownPeopleLabels() is recomputed on each NSComboBoxDataSource callback (numberOfItems, objectValueForItemAt, indexOfItemWithStringValue), so the same sort runs 3+ times per keystroke. At N=2-20 speakers this is negligible, but a per-edit-cycle cache would be a cleanliness win if the list ever grows.

What looks good:

• The completedLabel logic correctly requires exactly one unambiguous prefix match before completing — no surprising auto-completions when multiple names share a prefix.
• Stateless enum with pure static functions is the right design here: no threading concerns, trivially testable, and the @MainActor isolation stays where it belongs (in SpeakerRowView).
• Good test coverage: unique prefix match, ambiguous prefix refusal, and duplicate-name disambiguation all exercised.
• The match ranking (exact → prefix → word-start → label-prefix → substring → label-substring) with call-count tiebreaking is well-thought-out and produces intuitive sort order.
• The extraction is faithful — the logic in SpeakerNameSelectionPolicy matches the old inline implementations with no silent behavioral drift (aside from the diacritic folding note above).

---

Generated by Claude Code