fix: force tool calls for drifted intents and harden tool-use prompt

Small local models (qwen2.5-vl-3b, gemma-4-e4b) narrate actions ('Sure,
I'll pause the music') without emitting the tool call once the full
system prompt and 26-tool set are in context. Measured: qwen emitted
tools in 2/8 realistic intents, gemma 5/8; single-tool tests pass 100%,
so this is context-induced drift, not plumbing.

- agent/intent_router.py: deterministic forced-call router for
  high-frequency intents — media controls, now_playing/list_library,
  'remember that ...' memory saves, explicit YouTube Music searches,
  and personal-fact recall (memory_recall/search_notes/skill_recall,
  excluded for weather/time questions). Fires only when the tool is
  available; guards against how-do-I questions.
- agent.py: generalize _ForcedToolCallStream to any tool; route through
  intent_router after the dispatch check. System prompt adds a hard
  'call, don't narrate' rule and reconditions 'answer directly'.
- tests/test_intent_router.py: 30 unit tests pinning router behavior.

Verified against the live LLM with the full fidelity harness: combined
forced-or-emitted coverage 8/8 on qwen2.5-vl-3b (was 2/8) and 8/8 on
gemma-4-e4b (was 5/8). Full suite: 79 passed.

agent/agent.py

@@ -127,11 +127,19 @@ SYSTEM_PROMPT = textwrap.dedent("""\
       no markdown, and no formatting except the Azure delivery markers below.
     - Spell out numbers when it sounds more natural ("twenty twenty-six"
       instead of "2026").
-    - Make reasonable assumptions and answer directly. Ask a follow-up only
-      when it is genuinely impossible to proceed without one.
+    - Make reasonable assumptions and answer directly once any relevant tool
+      call has been made. Ask a follow-up only when it is genuinely
+      impossible to proceed without one.
     - Be warm and direct. Don't be sycophantic or overly formal.
     - If you don't know something, say so briefly and move on.
 
+    # Tools: call them, don't narrate them
+    When a request matches one of your tools, you MUST emit that tool call
+    in this reply. Never only say "I'll pause the music", "let me check", or
+    "I'm saving that now" without the tool call — narrating an action is a
+    failure because nothing actually happens. After the tool result comes
+    back, confirm briefly in your normal voice.
+
     # Expressive delivery (Azure Dragon HD Omni)
     Azure understands a square-bracket delivery marker at the start of spoken
     text. Use zero to two markers per response, only when they genuinely improve
@@ -223,12 +231,14 @@ SYSTEM_PROMPT = textwrap.dedent("""\
 """)
 
 
-# ── Background-task routing ─────────────────────────────────────────────────
+# ── Intent routing ──────────────────────────────────────────────────────────
 
 try:
     from agent.task_router import should_force_background_task
+    from agent.intent_router import detect_forced_call
 except ImportError:
     from task_router import should_force_background_task
+    from intent_router import detect_forced_call
 
 
 # ── OpenAI-compatible LLM with reasoning_content fallback ───────────────────
@@ -320,23 +330,39 @@ class _ReasoningFallbackWrapper:
 
 
 class _ForcedToolCallStream(lk_llm.LLMStream):
-    """Emit a dispatch_task call without asking the small LLM to route it."""
+    """Emit a tool call without asking the small LLM to produce it.
 
-    def __init__(self, llm, *, chat_ctx, tools, conn_options, description: str):
-        self._description = description
+    Two callers: the task router (dispatch_task for background work) and the
+    intent router (deterministic intents like "pause the music" or "remember
+    that ..."), where small models otherwise narrate the action without ever
+    emitting the call.
+    """
+
+    def __init__(
+        self,
+        llm,
+        *,
+        chat_ctx,
+        tools,
+        conn_options,
+        tool_name: str = "dispatch_task",
+        arguments: dict | None = None,
+    ):
+        self._tool_name = tool_name
+        self._arguments = arguments if arguments is not None else {}
         super().__init__(llm, chat_ctx=chat_ctx, tools=tools, conn_options=conn_options)
 
     async def _run(self) -> None:
         self._event_ch.send_nowait(
             lk_llm.ChatChunk(
-                id=f"forced-dispatch-{uuid.uuid4().hex}",
+                id=f"forced-{uuid.uuid4().hex}",
                 delta=lk_llm.ChoiceDelta(
                     role="assistant",
                     tool_calls=[
                         lk_llm.FunctionToolCall(
-                            name="dispatch_task",
-                            arguments=json.dumps({"description": self._description}),
-                            call_id=f"dispatch-{uuid.uuid4().hex}",
+                            name=self._tool_name,
+                            arguments=json.dumps(self._arguments),
+                            call_id=f"forced-{uuid.uuid4().hex}",
                         )
                     ],
                 ),
@@ -415,9 +441,11 @@ class HopeLLM(openai.LLM):
             )
             chat_ctx = self._compact_context(chat_ctx, conn_options)
 
-        # Force dispatch for clear background-work requests. Without this,
-        # small local models often say "I'll get on that" but emit no tool
-        # call at all.
+        # Force calls for intents small models drift on. Without this they
+        # say "I'll pause the music" or "I'll get on that" but emit no tool
+        # call at all. Task router handles background work; the intent router
+        # handles deterministic in-conversation intents (media controls,
+        # memory saves, music searches).
         items = list(chat_ctx.items)
         latest = items[-1] if items else None
         latest_role = getattr(latest, "role", None)
@@ -432,7 +460,24 @@ class HopeLLM(openai.LLM):
                 chat_ctx=chat_ctx,
                 tools=tools or [],
                 conn_options=conn_options,
-                description=latest_text,
+                tool_name="dispatch_task",
+                arguments={"description": latest_text},
+            )
+
+        forced = detect_forced_call(latest_text, tools or []) if latest_text else None
+        if forced:
+            tool_name, arguments = forced
+            logger.info(
+                "Routing request to forced %s(%s): %.120s",
+                tool_name, arguments, latest_text,
+            )
+            return _ForcedToolCallStream(
+                self,
+                chat_ctx=chat_ctx,
+                tools=tools or [],
+                conn_options=conn_options,
+                tool_name=tool_name,
+                arguments=arguments,
             )
 
         stream = super().chat(