Replace presecond speech profile trick with speaker embedding for user identification

[beastoin]

Problem

The backend listen pipeline uses a "presecond speech profile trick" to identify the user's voice: it prepends 10s of the user's speech profile audio + 5s padding before the actual stream, then assumes Deepgram's speaker == 0 is the user. This is:

1. Redundant — speaker embedding matching (_match_speaker_embedding() in transcribe.py:1809-1980) already exists and can identify the user by voice biometrics
2. Adds 15s startup delay — SPEECH_PROFILE_STABILIZE_DELAY = 15 seconds before actual transcription begins
3. Fragile — assumes the speech profile audio will cause Deepgram to assign speaker_id=0 to the user, which isn't guaranteed
4. Creates dual sockets — opens deepgram_socket (with preseconds=15) AND deepgram_profile_socket (preseconds=0), doubling Deepgram API costs per session

Current Architecture

The Presecond Trick (to be removed)

User speech profile (10s WAV) → 5s padding → actual audio stream
                                              ↓
                              Deepgram sees speaker_id=0 as "the user"
                              Words with start < 15s are filtered out

Key code locations:

• streaming.py:20-23 — Constants: SPEECH_PROFILE_FIXED_DURATION=10, PADDING=5, STABILIZE_DELAY=15
• streaming.py:347-350 — is_user = True if word.speaker == 0 and preseconds > 0
• transcribe.py:970-1150 — Dual socket creation, profile audio sending, stabilization wait
• transcribe.py:1117-1146 — send_initial_file_path() sends profile audio to Deepgram

Speaker Embedding (already exists, should replace)

User's stored embedding (512-dim vector)
    ↓
Extract embedding from live audio segment → cosine distance comparison
    ↓
If distance < 0.45 → same speaker (user identified)

Key code locations:

• utils/stt/speaker_embedding.py — extract_embedding_from_bytes(), compare_embeddings(), is_same_speaker()
• transcribe.py:1809-1980 — _match_speaker_embedding() already matches speakers in real-time
• database/users.py:240-264 — Speaker embedding stored per person in Firestore

Proposed Solution

1. Extract user's speech profile embedding at session start (or use pre-computed embedding from Firestore)
2. When Deepgram assigns speaker IDs, extract a short audio segment for each new speaker_id
3. Compare extracted embedding against user's profile embedding using existing is_same_speaker() (threshold 0.45)
4. If match → mark as user (same as current is_user=True behavior)
5. Remove presecond trick: no more profile audio prepending, no dual sockets, no 15s delay

Benefits

• Eliminates 15s startup delay — transcription begins immediately
• Saves Deepgram API costs — single socket instead of dual
• More accurate — voice biometric matching vs fragile speaker_id=0 assumption
• Simpler code — removes ~100 lines of presecond handling

Acceptance Criteria

• User identification uses speaker embedding comparison, not presecond trick
• Speech profile audio is NOT prepended to the stream
• Only one Deepgram socket opened per session (not two)
• No 15s startup delay — transcription begins immediately
• is_user flag is set correctly based on embedding match
• Existing speaker embedding infrastructure reused (no new ML models)
• All existing tests pass
• Backward compatible — users without speech profiles still work (just no user identification)