Add audio input support for HTTP Gateway and CLI (Phase 2)

[yacosta738]

Context

Follow-up from #246 (audio input support Phase 1). Phase 1 delivered core transcription infrastructure + Telegram channel. Phase 2 extends audio input to the remaining two entry points defined in the PRD.

Scope

HTTP Gateway

• New POST /web/chat/audio endpoint accepting multipart/form-data
• Fields: audio (file), session_id (optional), language (optional)
• Body limit increased to 25 MiB for this endpoint only
• Validate file, stage, transcribe, dispatch text through existing path
• Return transcription + agent response

CLI

• /audio <path> command for local file transcription
• Read local file, validate format/size/duration
• Stage as StagedAudio, transcribe, inject text

Optional

• whisper-rs embedded transcription behind --features audio-transcription (zero external dependency)
• Model auto-download tooling

Acceptance Criteria

• An integration can send audio through the HTTP Gateway and receive a normal agent response based on the transcription
• A user can pass an audio file through the CLI and get the same conversational behavior
• All supported formats work (OGG/Opus, MP3, WAV, M4A)
• Error handling matches the 6 error types from PRD
• Existing text and image flows unaffected

References

• PRD: GitHub Add audio input support for agents (Telegram, HTTP Gateway, CLI) #246 sections 7, 10 (FR7), 16 (acceptance criteria)
• Proposal: openspec/changes/archive/2026-04-03-audio-input-support/proposal.md (Phase 2 section)
• Design: openspec/changes/archive/2026-04-03-audio-input-support/design.md
• Spec: openspec/specs/audio-input/spec.md

[linear]

DALLAY-248