Cleanup personalization: optionally bias Apple Speech transcription with the same domain terms

[samkeen]

Follow-up from the cleanup-personalization feature (CHANGE_LOG 2026-06-15).

Context

We shipped a "Cleanup personalization" group in the Tuning sheet: two free-text fields (Domains, Terms & acronyms) persisted in Tunings.cleanup (CleanupSettings) and surfaced to the cleanup LLM as CleanupPersonalization via CleanupPrompt.system(personalization:). Scope was deliberately kept cleanup-only.

Idea

The same Terms & acronyms the user lists for cleanup are exactly the kind of names/jargon that Apple Speech's contextual biasing (AppleSpeechOptions.contextualStrings) exists to reduce misrecognitions of. Feeding them to the transcriber too would fix domain words at the source, before cleanup ever runs — "list once, help both stages."

Why it's deferred (the tension to resolve)

• There is already a separate Contextual biasing field on AppleSpeechSettingsSection (tunings.apple.contextualStringsText). Two overlapping inputs would confuse — we'd need to decide: merge them, derive one from the other, or keep both and union at the boundary.
• Contextual biasing is Apple-only; the cleanup terms are engine-independent. Coupling them blurs that line.
• Cleanup terms are free prose ("E2B is a Gemma variant"); contextual strings want short discrete tokens. Some parsing/extraction would be needed.

Sketch

• Option A: at the transcriptionOptions boundary for .apple, union apple.contextualStrings with the comma/term-split of cleanup.terms.
• Option B: collapse the two UI fields into one shared "Terms & acronyms" that feeds both transcription biasing and cleanup, and drop the separate Contextual biasing field.

Decide the UX first (one field vs two), then wire. No behavior change until then.