fix(context): cap summary output at model limit; safe token upper bound

- Clamp the summary request's max_tokens to the model's output limit so a
  large STRIX_CONTEXT_SUMMARY_TOKENS can't get the request rejected (which
  left the overflowing session uncompacted). Applied consistently to the
  input-budget reservation and the request itself.
- Replace the tokenizer-unavailable fallback with the UTF-8 byte length, a
  guaranteed upper bound on tokens for byte-level BPE, so budget checks can
  never under-count dense history.

tests/test_compaction.py

@@ -190,6 +190,19 @@ def test_fit_to_tokens_truncates_oversized_text(monkeypatch: pytest.MonkeyPatch)
     assert compaction._fit_to_tokens("m", "short", 500) == "short"
 
 
+def test_summary_output_tokens_capped_at_model_limit(monkeypatch: pytest.MonkeyPatch) -> None:
+    context = ContextSettings()
+    monkeypatch.setattr(compaction, "load_settings", lambda: SimpleNamespace(context=context))
+
+    monkeypatch.setattr(compaction, "output_limit", lambda _m: 1_000)
+    context.summary_max_tokens = 4_096
+    # Configured allowance above the model cap is clamped down to the cap.
+    assert compaction._summary_output_tokens("m") == 1_000
+    # Below the cap, the configured value is used unchanged.
+    context.summary_max_tokens = 500
+    assert compaction._summary_output_tokens("m") == 500
+
+
 @pytest.mark.asyncio
 async def test_maybe_compact_bounds_summary_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
     # A tiny window with a huge head must not send an oversized summary request.

tests/test_context_budget.py

@@ -38,10 +38,11 @@ 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)
-    # Conservative ~3-chars/token estimate, rounded up, so budgets never
+    # UTF-8 byte length is a guaranteed upper bound on tokens, 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
+    assert context_budget.count_tokens("weird-model", "x" * 400) == 400
+    # Multibyte text: bytes (not chars) bound the token count.
+    assert context_budget.count_tokens("weird-model", "😀" * 10) == 40
 
 
 def test_count_tokens_empty_is_zero() -> None: