feat(stt): Use the dictation API for transcription and LLM cleanup

[aurpsis-aai]

What

Switches AssemblyAITranscriber from the raw Sync STT endpoint (sync.assemblyai.com, X-AAI-Model: u3-sync-pro) to the dictation API (dictation.assemblyai.com/transcribe), so one request per utterance now covers both transcription and the post-processing LLM cleanup step server-side.

How

• The JSON config part gains an empty llm block — per the dictation API contract, { "llm": {} } requests the service's default cleanup rewrite (remove disfluencies, fix punctuation) under its server-owned guardrail scaffold. The rewrite instruction deliberately stays server-side; nothing is duplicated in the client.
• The response now decodes text (verbatim), llm_response, and llm_error. The engine pastes llm_response and falls back to text when it's null — the rewrite is best-effort (5 s server budget), so llm_error is logged as a degradation, never surfaced as a failure (required client behavior per the API doc).
• X-AAI-Model is no longer sent — the dictation service pins the STT model when it forwards to sync.
• The request carries the API's documented 90 s client timeout.
• TranscriptionPrompt still rides along as config.prompt, unchanged — it steers transcription (contextual priming, key terms); cleanup is now the llm block's job. The dictation service forwards prompt/sample_rate/channels to sync as-is, and the audio geometry/limits (16 kHz S16LE, ~0.1–120 s) are unchanged, so MicCapture and SyncSTTLimits behavior is untouched.
• Docs (AGENTS.md, BLURTENGINE.md, README.md) updated, including the "no separate LLM cleanup pass" invariant, which is now "no client-side LLM cleanup pass — the rewrite rides the same single request".

Architecture invariants preserved

One POST per utterance (no second request for the cleanup), no streaming, no client-side LLM pass, no new dependencies, TranscriberProtocol shape unchanged.

Tests

• Happy path asserts the rewrite (not the verbatim transcript) is returned, still in a single round-trip.
• New fallback test: null/absent llm_response (with and without llm_error) degrades to text.
• New config test: llm is always present and always empty.
• Wire-contract test updated: raw key auth, no X-AAI-Model, 90 s timeout.

Verification

• swift build -Xswiftc -warnings-as-errors clean; swift format lint --strict clean; scripts/check.sh --portable green.
• swift test could not run locally (no Xcode toolchain on this machine — no Testing module); per AGENTS.md, CI's check.yml on macos-26 is the authority on green.

Notes for reviewers

• The dictation API returns 404 (not 401) for an invalid key; the existing error path already extracts the body's detail ("Invalid API key"), so the user still sees a meaningful message.
• Latency: the rewrite adds LLM time to the round-trip (bounded at 5 s server-side; typical short clips well under that). The warm-up path is unchanged and now pre-pools the dictation host connection.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR migrates AssemblyAITranscriber in BlurtEngine from the Sync STT endpoint to AssemblyAI’s dictation API, so each utterance is handled by a single request that returns both the verbatim transcript and (when available) a server-side LLM cleanup rewrite that the app pastes.

Changes:

• Switch STT transport to https://dictation.assemblyai.com/transcribe, remove X-AAI-Model, and set a 90s client timeout.
• Always request the default server-side rewrite via an empty config.llm object and fall back to text when llm_response is missing/null (logging llm_error as degradation).
• Update engine/app/docs/tests wording and coverage to reflect “dictation API + rewrite” behavior.

Reviewed changes

Copilot reviewed 18 out of 18 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
Tests/BlurtEngineTests/TranscriptionPromptTests.swift	Updates test descriptions to reflect transcription prompt usage (not Sync-specific).
Tests/BlurtEngineTests/DictationSessionTests.swift	Updates comments to reflect dictation API semantics and unchanged short-clip behavior.
Tests/BlurtEngineTests/AssemblyAITranscriberTests.swift	Adds/updates tests for dictation endpoint, rewrite preference, fallback behavior, header/timeout contract, and config llm presence.
Sources/BlurtEngine/STT/TranscriptionPrompt.swift	Updates docs to describe dictation API prompt usage.
Sources/BlurtEngine/STT/TranscriptionContext.swift	Updates docs to remove Sync-specific wording.
Sources/BlurtEngine/STT/TranscriberProtocol.swift	Updates docs to reflect dictation API behavior (single-shot result).
Sources/BlurtEngine/STT/SyncSTTLimits.swift	Clarifies that Sync model limits still apply behind dictation API forwarding.
Sources/BlurtEngine/STT/AssemblyAITranscriber.swift	Core change: dictation endpoint, 90s timeout, remove model header, request rewrite via llm: {}, decode llm_response/llm_error, fallback to verbatim text.
Sources/BlurtEngine/Pipeline/DictationSession+Pipeline.swift	Updates pipeline comments to reflect dictation API and “no client-side styling pass”.
Sources/BlurtEngine/Pipeline/DictationSession.swift	Updates comments around recording cap and connection warm-up for dictation host.
Sources/BlurtEngine/Config/KeyTermsStore.swift	Updates docs to refer to dictation request priming.
Sources/BlurtEngine/Audio/MicCaptureProtocol.swift	Updates docs to refer to dictation request upload format.
Sources/BlurtEngine/Audio/MicCapture.swift	Updates docs to refer to dictation API upload expectations.
README.md	Updates product overview and architecture summary to reflect dictation API + rewrite.
BLURTENGINE.md	Updates engine developer guide to reflect dictation API request/response and server-side rewrite behavior.
App/Blurt/Blurt/Overlay/OverlayView.swift	Updates UI copy comments to refer to dictation API latency.
App/Blurt/Blurt/DictationComposition.swift	Updates composition docs to refer to dictation transcriber.
AGENTS.md	Updates canonical repo/architecture notes to reflect dictation API + server-side rewrite contract.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 18 out of 18 changed files in this pull request and generated 1 comment.