fix(llm): correct context sizing, honor think requests, broaden delete

Three related fixes uncovered while benchmarking qwen3:14b against 8b:

- pick_num_ctx was only counting message content, missing the ~15K
  tokens of tool schemas. num_ctx=8192 was being selected while actual
  prompt_tokens hit 14K+, causing silent prompt truncation on every
  tool-using request. Now includes json.dumps(tools) in the estimate.
  KV cache priming in app.py and routes/settings.py also fetches tools
  so the primed num_ctx matches what real chat requests will use.

- _should_think's heuristic classifier was overriding explicit
  think=true requests from the frontend toggle and MCP, gating on
  message length and regex patterns. Now a pass-through — the caller
  is the source of truth. quick_capture hardcodes think=False since
  it's a fast classification path that was relying on the old gating.

- delete_note description only mentioned "note or task", so the model
  refused to call it for entries created by save_person / save_place /
  create_list. Description now explicitly lists all five note_types it
  handles.

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

src/fabledassistant/services/generation_task.py

@@ -37,62 +37,24 @@ _TOOL_CALL_MARKER = re.compile(r"^\s*\[TOOL_CALLS\]\s*", re.IGNORECASE)
 DB_FLUSH_INTERVAL = 5.0  # seconds between partial DB flushes
 
 # ---------------------------------------------------------------------------
-# Conditional thinking classifier
+# Thinking decision
 # ---------------------------------------------------------------------------
-
-# Patterns that force think=True even on short messages
-_THINK_FORCE = re.compile(
-    r"\b("
-    r"analyz|compar|explain\s+why|help\s+me\s+(think|plan|understand|figure\s+out)|"
-    r"step[- ]by[- ]step|debug|troubleshoot|diagnos|"
-    r"pros\s+and\s+cons|trade[- ]?off|"
-    r"architect|design\s+(a|the|my|this)|"
-    r"write\s+a\s+(detailed|long|comprehensive|full)|"
-    r"brainstorm|outline\s+(a|the|my)|"
-    r"what\s+(are|is)\s+the\s+(best|difference|relationship|impact|implication)|"
-    r"how\s+(do|does|should|would|can)\s+.{0,40}\s+work|"
-    r"why\s+(is|are|does|do|did|would|should)\b"
-    r")",
-    re.IGNORECASE,
-)
-
-# Patterns that force think=False regardless of message length
-_THINK_SKIP = re.compile(
-    r"^(hi|hey|hello|thanks|thank\s+you|ok|okay|got\s+it|sounds\s+good|"
-    r"great|perfect|sure|yes|no|yep|nope|nice|cool|awesome|"
-    r"what('s| is) \d|what time|how many|remind me|add (a |an )?(task|note|reminder)|"
-    r"create (a |an )?(task|note)|delete|update|mark .{0,30} (done|complete))\b",
-    re.IGNORECASE,
-)
-
-_WORD_COUNT_THRESHOLD = 60  # messages over this word count always use think=True
-_SHORT_MESSAGE_THRESHOLD = 12  # messages under this always use think=False
+#
+# The `think` flag from the frontend / MCP is taken at face value: if the
+# caller asked for thinking, they get thinking. No heuristic gating.
+#
+# Models that don't support extended reasoning (e.g. llama3, mistral) simply
+# ignore the `think` parameter in the Ollama chat request, so this is safe to
+# pass unconditionally across the full model zoo.
 
 
 def _should_think(user_content: str, think_requested: bool) -> bool:
     """Return whether extended thinking should be used for this request.
 
-    If the caller didn't request thinking, we never enable it. If they did,
-    we check whether the message is complex enough to warrant the overhead.
+    Honors the caller's request directly — no message-complexity classifier.
+    The frontend toggle / MCP `think` parameter is the source of truth.
     """
-    if not think_requested:
-        return False
-
-    text = user_content.strip()
-    word_count = len(text.split())
-
-    if word_count <= _SHORT_MESSAGE_THRESHOLD:
-        return False
-    if _THINK_SKIP.match(text):
-        return False
-    if word_count >= _WORD_COUNT_THRESHOLD:
-        return True
-    if "```" in text:
-        return True
-    if _THINK_FORCE.search(text):
-        return True
-
-    return False
+    return bool(think_requested)
 
 
 # Human-readable labels for each tool, shown in the status indicator
@@ -273,9 +235,10 @@ async def run_generation(
         voice_speech_style=voice_speech_style,
     )
 
-    # Pick the smallest context tier that fits the current messages.
+    # Pick the smallest context tier that fits the current messages AND the
+    # tool schemas (which can be 6-10K tokens on their own with ~40 tools).
     # Using the minimum needed tier reduces KV cache size and speeds up prefill.
-    num_ctx = pick_num_ctx(messages)
+    num_ctx = pick_num_ctx(messages, tools=tools)
     logger.debug("Adaptive num_ctx=%d for conv %d", num_ctx, conv_id)
 
     # Emit context event