feat(profile): /api/profile + eligibility filter + inference framework (ADR-0014 steps 4-6)

Step 4 — /api/profile read-through API:
  GET  /api/profile          → { user, prefs, consents, contexts }
  PATCH /api/profile/prefs/:scope  upsert user_preferences (source='user')
  PATCH /api/profile/consents      grant / revoke consent keys
  PATCH /api/profile/contexts      create / activate / deactivate contexts
  Legacy consentGiven bit folded in as data:core fallback.

Step 5 — registry-driven eligibility filter:
  fetchRegistry() exported from agent-registry.ts.
  profile/eligibility.ts: getEligibleAgentIds(userId) — filters by required
  consents, silenced_in_contexts, and user_preferences[enabled=false].
  fetchOrchestratorTip filters agent_outputs to eligible set before calling
  ml/serving /recommend. Fail-closed: registry unavailable → empty set.

Step 6 — shared context-inference framework (#111) + time-of-day proof (#112):
  ml/agents/inference/: UserHistory, FeedbackEvent, run_inference().
  Framework: cold-start, min_history gating, error fallback, structured logs.
  TimeOfDayAgent v1.1.0: inferred_params=[preferred_hour]; also reads
  quiet_start/quiet_end from agent_prefs. agent_prefs injected by TS caller.
  AgentInput gains agent_prefs field.
  ml/serving: POST /agents/{agent_id}/infer endpoint.
  agent-outputs.ts computeAndStore: loads prefs before compute, calls /infer
  after, persists results (source='inferred'); user overrides never touched.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ml/serving/main.py

@@ -3,6 +3,7 @@ oO ML Serving — multi-agent orchestrator (ADR-0013).
 
 Contract:
   POST /agents/{agent_id}/compute  run a sub-agent, return prompt snippet
+  POST /agents/{agent_id}/infer    run inference framework for a user, return inferred prefs
   POST /recommend                  orchestrate agent snippets → one tip via LiteLLM
   POST /generate                   LLM tip candidates (legacy; kept for bench/eval)
   GET  /health                     { ok, agents: [...] }
@@ -38,7 +39,8 @@ if _repo_root not in sys.path:
     sys.path.insert(0, _repo_root)
 
 from ml.agents.base import AgentInput  # noqa: E402
-from ml.agents.registry import get_agent, all_agents, all_manifests  # noqa: E402
+from ml.agents.registry import get_agent, all_agents, all_manifests, get_manifest  # noqa: E402
+from ml.agents.inference import run_inference, FeedbackEvent, UserHistory  # noqa: E402
 
 logging_config.configure()
 
@@ -123,6 +125,8 @@ class AgentComputeRequest(BaseModel):
     profile: dict[str, Optional[float]] = {}
     feedback_history: list[dict] = []
     now_iso: Optional[str] = None  # ISO 8601; defaults to utcnow
+    # Per-agent prefs from user_preferences (merged: user source overrides inferred).
+    agent_prefs: dict = {}
 
 
 class AgentComputeResponse(BaseModel):
@@ -135,6 +139,18 @@ class AgentComputeResponse(BaseModel):
     agent_version: str
 
 
+class AgentInferRequest(BaseModel):
+    user_id: str
+    feedback_history: list[dict] = []   # [{action, dwell_ms, created_at}, …]
+
+
+class AgentInferResponse(BaseModel):
+    user_id: str
+    agent_id: str
+    # {key: inferred_value} — caller persists to user_preferences with source='inferred'
+    inferred_prefs: dict
+
+
 class AgentOutputSnippet(BaseModel):
     agent_id: str
     prompt_text: str
@@ -225,6 +241,7 @@ async def compute_agent(agent_id: str, req: AgentComputeRequest) -> AgentCompute
         profile=req.profile,
         feedback_history=req.feedback_history,
         now=now,
+        agent_prefs=req.agent_prefs,
     )
     try:
         output = agent.compute(inp)
@@ -244,6 +261,46 @@ async def compute_agent(agent_id: str, req: AgentComputeRequest) -> AgentCompute
     )
 
 
+@app.post("/agents/{agent_id}/infer", response_model=AgentInferResponse)
+async def infer_agent(agent_id: str, req: AgentInferRequest) -> AgentInferResponse:
+    """Run the inference framework for one agent and return inferred preference values.
+
+    The caller (TS agent-outputs.ts) persists results to user_preferences
+    with source='inferred', skipping keys where source='user' already exists.
+    """
+    try:
+        manifest = get_manifest(agent_id)
+    except KeyError:
+        raise HTTPException(status_code=404, detail=f"Unknown agent: {agent_id!r}")
+
+    if not manifest.inferred_params:
+        return AgentInferResponse(user_id=req.user_id, agent_id=agent_id, inferred_prefs={})
+
+    events = [
+        FeedbackEvent(
+            action=e.get("action", ""),
+            dwell_ms=e.get("dwell_ms"),
+            created_at=e.get("created_at", ""),
+        )
+        for e in req.feedback_history
+    ]
+    history = UserHistory(user_id=req.user_id, events=events)
+
+    t0 = __import__("time").monotonic()
+    inferred = run_inference(manifest, history)
+    latency_ms = round((__import__("time").monotonic() - t0) * 1000, 1)
+
+    log.info(
+        "inference_run",
+        agent_id=agent_id,
+        user_id=req.user_id,
+        n_params=len(inferred),
+        history_len=len(events),
+        latency_ms=latency_ms,
+    )
+    return AgentInferResponse(user_id=req.user_id, agent_id=agent_id, inferred_prefs=inferred)
+
+
 @app.post("/recommend", response_model=RecommendResponse)
 async def recommend(req: RecommendRequest) -> RecommendResponse:
     """Orchestrator: combine pre-computed agent outputs into one tip via LLM.