feat: on-device llm rewrite pass for dictated takes

[joshuagruenstein]

What

Optional post-dictation polish using Apple's on-device Foundation model (Apple Intelligence, macOS 26+), configured via a new literate [rewrite] section: templated switches (removeFillers, fixPunctuation, fixGrammar, correctVocabulary, technicalFormatting) plus one freeform customInstructions. Off by default.

Design

• Mechanical vs judgment split: um/uh/er stripping is a deterministic regex (works even without Apple Intelligence); the model only does punctuation, dedup, vocabulary-aware mis-transcription repair, and symbol-speak.
• Transcript is data, not conversation: sent wrapped in <transcript> tags with instructions that questions/requests inside are things the speaker said. Without this, the 3B model answers meta-dictation instead of editing it.
• Acceptability gate: rejects empty/ballooned output, low word retention (answers/refusals), and silent profanity censorship (every profane token in must come out, same count — the model's safety training otherwise swaps swears for tame verbs). Rejection degrades to the filler-stripped text; the user's words are never at risk.
• Apply: single prefix-cut keystroke edit (delete back to the first changed character, retype the rest). Backspaces physically cannot preserve a common suffix, so no other diff shape is valid — pinned by regression tests that simulate field semantics, including grapheme cases.
• Cancellation: new squeeze, erase, send, or app switch invalidates a pending rewrite (generation counter + take-identity + frontmost-app checks). Blue menu bar dot while the model runs.
• Stage-by-stage logging (in → stripped → model out → gate verdict → applied edit) for future debugging.

Prompt was tuned across four eval rounds against 50 real dictations via the new --rewrite-eval harness (runs the production pipeline against a sample file): final round has zero gate rejections, ~1.7s average latency.

Also carries two live-verified UX fixes: sequenced permission prompts (mic → accessibility, so the second dialog isn't buried) and no OK button on in-progress cards.

Testing

• 283 assertions green (swift run tingle-tests), including new suites for filler strip, gate, prompt assembly, profanity backstop, and the edit/apply path
• Payload + version tooling tests green
• Live-tested by Josh on the dev build (rewrite quality, blue dot, permission sequencing, progress cards)

🤖 Generated with Claude Code