fix(context): over-estimate tokens when no tokenizer is available

The chars/4 fallback under-counts dense text (code, base64, CJK), which
could let a summary request be packed past the real context window and get
rejected. Use a conservative ~3-chars/token estimate instead so budget
checks never under-count.

strix/llm/context_budget.py

@@ -68,10 +68,18 @@ def output_limit(model: str) -> int:
 
 
 def count_tokens(model: str, text: str) -> int:
-    """Token count for ``text`` under ``model`` (chars/4 fallback)."""
+    """Token count for ``text`` under ``model``.
+
+    LiteLLM's counter handles known tokenizers (and defaults to a tiktoken
+    encoding otherwise). If it still can't count, fall back to a *conservative*
+    estimate: token density varies, and dense text (code, base64, CJK) can run
+    well under 4 chars/token, so we assume ~3 to over-estimate rather than
+    under-estimate — an under-estimate would let a summary request be packed
+    past the real context window and get rejected.
+    """
     if not text:
         return 0
     try:
         return int(litellm.token_counter(model=_lookup_key(model), text=text))
     except Exception:  # noqa: BLE001 - tokenizer may be unavailable for some models.
-        return len(text) // 4
+        return -(-len(text) // 3)

tests/test_context_budget.py

@@ -38,8 +38,10 @@ def test_count_tokens_fallback_on_error(monkeypatch: pytest.MonkeyPatch) -> None
         raise RuntimeError("no tokenizer")
 
     monkeypatch.setattr("strix.llm.context_budget.litellm.token_counter", _raise)
-    text = "x" * 400
-    assert context_budget.count_tokens("weird-model", text) == 100
+    # Conservative ~3-chars/token estimate, rounded up, so budgets never
+    # under-count when no tokenizer is available.
+    assert context_budget.count_tokens("weird-model", "x" * 400) == 134
+    assert context_budget.count_tokens("weird-model", "x") == 1
 
 
 def test_count_tokens_empty_is_zero() -> None: