fix(desktop): dedupe auto-spoken replies by content fingerprint

[lilShawtty-byte]

Problem

Auto-speak (voice.auto_tts) sometimes reads the same reply aloud twice.

Root cause: spoken-reply dedupe tracked the last spoken message by its id, but message ids are NOT stable across the post-turn hydrate:

1. During streaming, the bubble carries a runtime id (assistant-stream-<ts>-<seq>)
2. On message.complete, completeAssistantMessage settles that bubble and auto-speak reads it (first playback)
3. The same handler then calls hydrateFromStoredSession → toChatMessages rebuilds the list with persisted ids (<timestamp>-<index>-<role>)
4. `` emits again → pendingReply sees a 'new' id ≠ `lastSpokenIdRef` → plays the same text a second time

Both existing dedupe layers fail: lastSpokenIdRef (id rewritten by hydrate) and ownsAmbientCue (1s window, the two triggers are a network round-trip apart).

Fix

• lib/chat-messages.ts: add speechFingerprint() (trimmed text) and latestSpokenReply() — an auto-speak selector that dedupes on content, which survives the hydrate id rewrite, and returns the fingerprint as the ambient-cue key.
• use-composer-voice.ts: auto-speak path (pendingResponse / consumePendingResponse) uses the fingerprint ref. The voice-conversation path keeps id semantics — its live speech session binds to a message id for the turn and depends on it staying stable while streaming.

Verification

• vitest run src/lib/chat-messages.test.ts → 57/57 (5 new tests incl. the streaming→hydrated id-rewrite case)
• vitest run src/app/chat/composer/hooks/use-voice-conversation*.test.tsx → 11/11 (voice-conversation path unaffected)
• tsc --noEmit: no new errors (11 pre-existing errors in user-edit-composer.tsx, unrelated @assistant-ui version drift)

[teknium1]

Thanks for tracing the runtime-to-hydrated id transition; current main does have the id-only selector at apps/desktop/src/app/chat/composer/hooks/use-composer-voice.ts:75-109, and hydration replaces the message list at apps/desktop/src/app/contrib/wiring.tsx:321-327.

Problems

• apps/desktop/src/lib/chat-messages.ts:211 makes trimmed text the entire persistent spoken identity. A later distinct reply with the same content as the preceding reply is therefore returned as null and never read aloud. useAutoSpeakReplies is intended to read each completed assistant turn (apps/desktop/src/app/chat/composer/hooks/use-auto-speak-replies.ts:29-34).

Suggested changes

• Rework the hydration dedupe to retain or reconcile a per-turn identity rather than using content alone, and add a regression case for two distinct same-text replies. The current new-reply test uses different text, so it does not exercise this collision.

Automated hermes-sweeper review.

[teknium1]

This permanently treats equal text as the same reply. A later, distinct turn returning the same text (for example, two separate Done. replies) will be suppressed rather than spoken. Please retain or reconcile a per-turn identity across hydration and add a same-text, distinct-turn regression case.

[OutThisLife]

Superseded by #89815.

That salvage keeps the stream-id rewrite fix, rebased onto current main, and keys spoken state on the assistant-turn ordinal instead of reply text so two Done. turns in one session still both speak. Manual Read Aloud marks the same book after a successful play.

You're credited via Co-authored-by. Thanks for tracing the live-to-durable id swap.