fix(model_metadata): drop stale ≤256,000 cache entries for Grok-4.3

The ``grok-4.3`` (1M context) catalog entry was added on 2026-05-15
(ce0e189d3).  Between 2026-04-10 (when ``grok-4`` at 256,000 was first
added by b57769718) and 2026-05-15, grok-4.3 slugs resolved via the
generic ``grok-4`` substring catch-all and that 256,000 value was
persisted to context_length_cache.yaml.  Users who first queried
grok-4.3 in that 35-day window are stuck at 256K forever — the cache
is read at step 1 before the hardcoded defaults in step 8, so the
correct 1M entry is never reached.

Mirror the existing Kimi/Codex/MiniMax-M3 stale-cache guards: add
_model_name_suggests_grok_4_3() and an elif branch that drops any
cached value ≤ 256,000 for a grok-4.3 slug so the next lookup falls
through to the 1M hardcoded default.

Adds 4 regression tests: helper unit test, stale-drop-and-re-resolve,
correct-cache-preserved, and no-clobber for plain grok-4 (256K correct).

agent/model_metadata.py

@@ -1140,6 +1140,18 @@ def _model_name_suggests_minimax_m3(model: str) -> bool:
     return "minimax-m3" in model.lower()
 
 
+def _model_name_suggests_grok_4_3(model: str) -> bool:
+    """Return True if the model name looks like a Grok 4.3 variant.
+
+    Catches ``grok-4.3``, ``grok-4.3-latest``, and similar slugs.
+    Used as a guard against stale cache entries seeded by pre-catalog builds
+    that resolved grok-4.3 via the generic ``grok-4`` catch-all (256,000)
+    before the ``grok-4.3`` (1M) entry was added to DEFAULT_CONTEXT_LENGTHS
+    on 2026-05-15.
+    """
+    return "grok-4.3" in model.lower()
+
+
 def _query_local_context_length(model: str, base_url: str, api_key: str = "") -> Optional[int]:
     """Query a local server for the model's context length."""
     import httpx
@@ -1564,6 +1576,19 @@ def get_model_context_length(
                     model, base_url, f"{cached:,}",
                 )
                 _invalidate_cached_context_length(model, base_url)
+            # Invalidate stale ≤256,000 cache entries for Grok-4.3.  The
+            # ``grok-4.3`` (1M) entry was added to DEFAULT_CONTEXT_LENGTHS on
+            # 2026-05-15; prior to that, grok-4.3 slugs resolved via the
+            # ``grok-4`` catch-all (256,000) and that value was persisted.
+            # grok-4.3 is 1M, so any sub-262K cached value is a pre-catalog
+            # leftover — drop it and fall through to the hardcoded default.
+            elif cached <= 256_000 and _model_name_suggests_grok_4_3(model):
+                logger.info(
+                    "Dropping stale Grok-4.3 cache entry %s@%s -> %s (pre-catalog value); "
+                    "re-resolving via hardcoded defaults",
+                    model, base_url, f"{cached:,}",
+                )
+                _invalidate_cached_context_length(model, base_url)
             # Nous Portal: the portal /v1/models endpoint is authoritative.
             # Bypass the persistent cache so step 5b can always reconcile
             # against it — this corrects pre-fix entries seeded from the

tests/agent/test_model_metadata.py

@@ -1247,3 +1247,59 @@ class TestContextLengthCache:
         with patch("agent.model_metadata._get_context_cache_path", return_value=cache_file):
             save_context_length(model, url, 200000)
             assert get_cached_context_length(model, url) == 200000
+
+
+class TestGrok43StaleCacheGuard:
+    """Pre-catalog builds resolved grok-4.3 via the generic 'grok-4' catch-all
+    (256,000) and persisted it before the 'grok-4.3' (1M) catalog entry was
+    added on 2026-05-15.  The step-1 cache guard must drop that stale value
+    and re-resolve to 1M, while leaving correct grok-4 entries (256,000)
+    untouched.
+    """
+
+    def test_suggests_grok_4_3(self):
+        from agent.model_metadata import _model_name_suggests_grok_4_3
+        assert _model_name_suggests_grok_4_3("grok-4.3")
+        assert _model_name_suggests_grok_4_3("grok-4.3-latest")
+        assert _model_name_suggests_grok_4_3("xai/grok-4.3")
+        assert not _model_name_suggests_grok_4_3("grok-4")
+        assert not _model_name_suggests_grok_4_3("grok-4-fast")
+        assert not _model_name_suggests_grok_4_3("grok-4.20")
+
+    def test_stale_grok_4_3_dropped_and_reresolves_to_1m(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        import importlib
+        import agent.model_metadata as mm
+        importlib.reload(mm)
+        base = "https://api.x.ai/v1"
+        mm.save_context_length("grok-4.3", base, 256_000)
+        ctx = mm.get_model_context_length(
+            "grok-4.3", base_url=base, api_key="", provider="xai"
+        )
+        assert ctx == 1_000_000
+
+    def test_correct_grok_4_3_cache_preserved(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        import importlib
+        import agent.model_metadata as mm
+        importlib.reload(mm)
+        base = "https://api.x.ai/v1"
+        mm.save_context_length("grok-4.3", base, 1_000_000)
+        ctx = mm.get_model_context_length(
+            "grok-4.3", base_url=base, api_key="", provider="xai"
+        )
+        assert ctx == 1_000_000
+
+    def test_grok_4_not_clobbered(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        import importlib
+        import agent.model_metadata as mm
+        importlib.reload(mm)
+        base = "https://api.x.ai/v1"
+        # 256,000 is the CORRECT value for plain grok-4 — guard must not touch it.
+        for slug in ("grok-4", "grok-4-0709"):
+            mm.save_context_length(slug, base, 256_000)
+            ctx = mm.get_model_context_length(
+                slug, base_url=base, api_key="", provider="xai"
+            )
+            assert ctx == 256_000, f"{slug} should stay 256000, got {ctx}"