fix(slots): warn on container config drift

src/hal0/slots/manager.py

@@ -29,6 +29,7 @@
 import asyncio
 import contextlib
 import logging
+import os
 import re
 import shutil
 import time
@@ -111,6 +112,8 @@
 # /health) is demoted to ERROR — but only after this many CONSECUTIVE failures,
 # so a single transient blip doesn't trigger a disruptive model reload.
 _HEALTH_FAIL_STRIKES: int = 2
+# Must match the exact flag spellings emitted by ContainerProvider's llama
+# launch renderer; drift checks compare argv text, not llama-server aliases.
 _CONFIG_DRIFT_KEYS: tuple[str, ...] = ("--ctx-size", "--model", "--alias", "-b", "-ub")
 
 # Idle-monitor defaults. A READY slot whose last activity is older than
@@ -1252,7 +1255,7 @@ async def compute_config_drift(
         diffs = [
             {"key": key, "running": running_flags.get(key), "rendered": rendered_flags.get(key)}
             for key in _CONFIG_DRIFT_KEYS
-            if running_flags.get(key) != rendered_flags.get(key)
+            if not _config_drift_values_equal(key, running_flags.get(key), rendered_flags.get(key))
         ]
         return {"drifted": bool(diffs), "diffs": diffs}
 
@@ -2567,6 +2570,12 @@ def _argv_values(argv: list[str], keys: tuple[str, ...]) -> dict[str, str | None
     return out
 
 
+def _config_drift_values_equal(key: str, running: str | None, rendered: str | None) -> bool:
+    if key == "--model" and running is not None and rendered is not None:
+        return os.path.realpath(running) == os.path.realpath(rendered)
+    return running == rendered
+
+
 def _normalize_ctx_key(cfg_dict: dict[str, Any]) -> None:
     """Fold the legacy ``[model].ctx_size`` alias into the canonical
     ``context_size`` (SlotConfig's field), in place (#585).

tests/providers/test_container.py

@@ -524,6 +524,28 @@ def test_expected_argv_uses_launch_plan_context_derive(self) -> None:
         assert argv is not None
         assert argv[argv.index("--ctx-size") + 1] == "32768"
 
+    def test_expected_argv_emits_config_drift_watched_flag_spellings(self) -> None:
+        """#863: drift watches exact argv spellings, so renderer renames must fail."""
+        provider = self._provider()
+        with patch(
+            "hal0.providers.container._resolve_profile",
+            return_value=_moe_profile(),
+        ):
+            argv = provider.expected_argv(
+                _slot_cfg(model={"default": "chadrock-35b-ace-saber", "context_size": 131072}),
+                _model_info(),
+            )
+
+        assert argv is not None
+        assert "--model" in argv
+        assert "--alias" in argv
+        assert "--ctx-size" in argv
+        assert "-b" in argv
+        assert "-ub" in argv
+        assert "-c" not in argv
+        assert "--batch-size" not in argv
+        assert "--ubatch-size" not in argv
+
 
 # ── load_sync / unload_sync systemd interaction ───────────────────────────────
 

tests/slots/test_manager.py

@@ -527,6 +527,41 @@ async def test_status_omits_config_drift_when_argv_matches(
     assert snap.metadata.get("config_drift") == {"drifted": False, "diffs": []}
 
 
+async def test_status_omits_config_drift_when_model_paths_resolve_to_same_file(
+    slot_root: Path,
+    tmp_path: Path,
+    container_stub: FakeContainerProvider,
+) -> None:
+    """#863: --model compares real paths so symlink/remount spelling is stable."""
+    real_dir = tmp_path / "models"
+    real_dir.mkdir()
+    model_path = real_dir / "qwen.gguf"
+    model_path.write_text("", encoding="utf-8")
+    link_dir = tmp_path / "model-link"
+    link_dir.symlink_to(real_dir, target_is_directory=True)
+
+    expected = [
+        "--host",
+        "0.0.0.0",
+        "--port",
+        "8081",
+        "--model",
+        str(model_path),
+        "--ctx-size",
+        "131072",
+    ]
+    running = list(expected)
+    running[running.index("--model") + 1] = str(link_dir / "qwen.gguf")
+    container_stub.expected_argv_by_slot["chat"] = expected
+    container_stub.running_argv_by_slot["chat"] = running
+
+    sm = SlotManager()
+    await sm.load("chat")
+    snap = await sm.status("chat", include_config_drift=True)
+
+    assert snap.metadata.get("config_drift") == {"drifted": False, "diffs": []}
+
+
 async def test_list_does_not_compute_config_drift_on_poll_path(
     slot_root: Path,
     container_stub: FakeContainerProvider,