fix(speaker): create a tenant on first use instead of seeding one at startup

Making users.id a Chronicle id broke the cold-start path, which CI caught and a
warm local database cannot: startup called ensure_admin_user ->
User(username="admin") with no id. That relied on integer autoincrement, so it
now inserts a NULL primary key, the service dies before serving /health, and
the integration job waits 600s for a container that will never come up.

A tenant is a Chronicle user id, so the service cannot invent one before a
caller names it. Nothing is seeded; the row is created on that user's first
enrolment, and get_or_create_user is keyed on the id rather than the username.

Three more places still assumed the integer:

- extract_user_id_from_speaker_id parsed int(speaker_id.split("_")[1]). It now
  reads speakers.user_id, which is the actual ownership record — a parse of the
  id convention is not, and it stops being true for an imported or
  differently-formatted speaker. This is the "do not infer the tenant from a
  SpeakerId" case from the audit.
- The websocket enhancement path ran int() over the tenant and fell back to
  None on failure, so every ObjectId silently became "no user" — which reads
  downstream as "search every gallery", the opposite of what it must mean.
- Six log lines formatted the tenant with %d, and the FAISS position map was
  typed Dict[int, Tuple[int, str]].

Cold-database regression tests added, since no warm database exercises this.

Author: AnkushMalaker

extras/speaker-recognition/src/simple_speaker_recognition/api/core/utils.py

@@ -105,24 +105,30 @@ def secure_temp_file(suffix: str = ".wav") -> tempfile._TemporaryFileWrapper:
     return tempfile.NamedTemporaryFile(delete=False, suffix=suffix)
 
 
-def extract_user_id_from_speaker_id(speaker_id: str) -> int:
-    """Extract user_id from speaker_id format: user_{user_id}_..."""
-    if not speaker_id.startswith("user_"):
-        raise HTTPException(
-            400,
-            f"Invalid speaker_id format. Expected 'user_{{user_id}}_...', got: {speaker_id}",
-        )
+def extract_user_id_from_speaker_id(speaker_id: str) -> str:
+    """Return the tenant that owns ``speaker_id``, read from the speaker row.
+
+    Deliberately a lookup rather than a parse. The id is *conventionally*
+    ``user_{user_id}_speaker_{hex}``, but that convention is not the ownership
+    record — the ``speakers.user_id`` column is, and it stays right when a speaker is
+    imported, renamed, or created by a client that formats ids differently. The old
+    ``int(speaker_id.split("_")[1])`` also stopped working the moment a tenant became
+    a Chronicle ObjectId, which is not an integer.
+    """
 
+    # Imported here to avoid a circular import: database.queries imports the API
+    # models that this utils module is itself imported by.
+    from simple_speaker_recognition.database import get_db_session
+    from simple_speaker_recognition.database.models import Speaker
+
+    db = get_db_session()
     try:
-        parts = speaker_id.split("_")
-        if len(parts) < 2:
-            raise ValueError("Not enough parts")
-        user_id = int(parts[1])
-        return user_id
-    except (ValueError, IndexError):
-        raise HTTPException(
-            400, f"Invalid speaker_id format. Cannot extract user_id from: {speaker_id}"
-        )
+        speaker = db.query(Speaker).filter(Speaker.id == speaker_id).first()
+        if speaker is None:
+            raise HTTPException(404, f"Speaker not found: {speaker_id}")
+        return str(speaker.user_id)
+    finally:
+        db.close()
 
 
 def validate_confidence(confidence: Any, context: str = "") -> float:

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/deepgram_wrapper.py

@@ -19,6 +19,7 @@
     Request,
     UploadFile,
 )
+
 from simple_speaker_recognition.api.core.utils import (
     safe_format_confidence,
     validate_confidence,

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/enrollment.py

@@ -10,6 +10,8 @@
 
 import numpy as np
 from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile
+from sqlalchemy import func
+
 from simple_speaker_recognition.api.core.utils import (
     extract_user_id_from_speaker_id,
     get_data_directory,
@@ -19,7 +21,6 @@
 from simple_speaker_recognition.database import get_db_session
 from simple_speaker_recognition.database.models import Speaker, SpeakerAudioSegment
 from simple_speaker_recognition.utils.audio_processing import get_audio_info
-from sqlalchemy import func
 
 # These will be imported from the main service.py when we integrate
 # from ..service import get_db, audio_backend, auth

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/enrollment_audit.py

@@ -19,6 +19,7 @@
 from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, UploadFile
 from fastapi.responses import FileResponse
 from pydantic import BaseModel, Field
+
 from simple_speaker_recognition.api.core.utils import secure_temp_file
 from simple_speaker_recognition.core.enrollment_audit import (
     compute_audit,

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/identification.py

@@ -13,6 +13,7 @@
 import numpy as np
 from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, UploadFile
 from pydantic import BaseModel
+
 from simple_speaker_recognition.api.core.utils import (
     safe_format_confidence,
     secure_temp_file,

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/speakers.py

@@ -10,6 +10,7 @@
 import numpy as np
 from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, UploadFile
 from fastapi.responses import FileResponse
+
 from simple_speaker_recognition.api.core.utils import extract_user_id_from_speaker_id
 from simple_speaker_recognition.constants import DEFAULT_SIMILARITY_THRESHOLD
 from simple_speaker_recognition.core.unified_speaker_db import UnifiedSpeakerDB

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/users.py

@@ -1,6 +1,7 @@
 """User management endpoints."""
 
 from fastapi import APIRouter
+
 from simple_speaker_recognition.core.models import UserRequest, UserResponse
 from simple_speaker_recognition.database import get_db_session
 from simple_speaker_recognition.database.queries import UserQueries
@@ -31,7 +32,7 @@ async def create_user(request: UserRequest):
     """Create or get existing user."""
     db = get_db_session()
     try:
-        user = UserQueries.get_or_create_user(db, request.username)
+        user = UserQueries.get_or_create_user(db, request.user_id, request.username)
         return UserResponse(
             id=user.id, username=user.username, created_at=user.created_at.isoformat()
         )

extras/speaker-recognition/src/simple_speaker_recognition/api/routers/websocket_wrapper.py

@@ -24,6 +24,7 @@
 )
 from pyannote.audio import Model
 from pyannote.audio.pipelines import VoiceActivityDetection
+
 from simple_speaker_recognition.api.core.utils import (
     safe_format_confidence,
     validate_confidence,
@@ -804,14 +805,10 @@ async def deepgram_proxy_websocket(
     except (ValueError, TypeError):
         confidence_threshold = DEFAULT_SIMILARITY_THRESHOLD
 
-    # Parse user_id
-    if enhancement_params["user_id"]:
-        try:
-            user_id = int(enhancement_params["user_id"])
-        except (ValueError, TypeError):
-            user_id = None
-    else:
-        user_id = None
+    # The tenant is an opaque Chronicle id, so it is taken as given rather than
+    # coerced. int() here silently turned every ObjectId into "no user", which reads
+    # downstream as "search every gallery".
+    user_id = enhancement_params["user_id"] or None
 
     log.info(
         f"Enhancement params - user_id: {user_id}, confidence_threshold: {confidence_threshold}"

extras/speaker-recognition/src/simple_speaker_recognition/api/service.py

@@ -16,6 +16,7 @@
 from omegaconf import OmegaConf
 from pydantic import Field
 from pydantic_settings import BaseSettings
+
 from simple_speaker_recognition.api.core.utils import get_data_directory
 from simple_speaker_recognition.constants import DEFAULT_SIMILARITY_THRESHOLD
 from simple_speaker_recognition.core.audio_backend import AudioBackend
@@ -105,7 +106,7 @@ class Settings(BaseSettings):
     # Backend API configuration for chunked processing
     # Loaded from root config.yml speaker_recognition section, can be overridden by env vars
     max_diarize_duration: int = Field(
-        default=60,
+        default=1200,
         description="Maximum audio duration (seconds) for single PyAnnote call",
     )
     diarize_chunk_overlap: float = Field(
@@ -132,7 +133,9 @@ def __init__(self, **kwargs):
             "max_diarize_duration" not in kwargs
             and "MAX_DIARIZE_DURATION" not in os.environ
         ):
-            kwargs["max_diarize_duration"] = root_config.get("max_diarize_duration", 60)
+            kwargs["max_diarize_duration"] = root_config.get(
+                "max_diarize_duration", 1200
+            )
 
         if (
             "diarize_chunk_overlap" not in kwargs
@@ -278,9 +281,8 @@ async def lifespan(app: FastAPI):
     auth.enrollment_audio_dir.mkdir(parents=True, exist_ok=True)
     log.info("Enrollment audio directory ready: %s", auth.enrollment_audio_dir)
 
-    # Ensure admin user exists
-    admin_user_id = speaker_db.ensure_admin_user()
-    log.info("Admin user ready ✔ – user_id=%s", admin_user_id)
+    # No tenant is seeded here. A tenant is a Chronicle user id, which only a caller
+    # can supply, so rows are created on first enrolment instead.
 
     # Start the CUDA self-heal watchdog (no-op on CPU).
     watchdog_task = asyncio.create_task(_cuda_watchdog())

extras/speaker-recognition/src/simple_speaker_recognition/core/enrollment_audit.py

@@ -23,6 +23,7 @@
 from typing import Optional
 
 import numpy as np
+
 from simple_speaker_recognition.database.models import (
     EnrollmentAuditDecision,
     Speaker,