chore: make dry-run LLM load total; drop -O-stripped assert (#3717)

Address PR review:
- dry-run LLM load now catches any store exception (filelock.TimeoutError from
  lock contention, OSError, validation errors) and records it as a diagnostic
  ("Could not load ...", distinct from "not found") instead of propagating and
  crashing the editor preview (#3719). Adds a regression test with a load that
  raises TimeoutError.
- replace `assert llm is not None` (stripped under python -O) with an explicit
  guard so a None LLM can never reach _build_openhands_settings.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Simon Rosenberg

openhands-sdk/openhands/sdk/profiles/resolver.py

@@ -331,10 +331,19 @@ def resolve_agent_profile_dry_run(
             llm = llm_store.load(profile.llm_profile_ref, cipher=cipher)
             diagnostics.llm_profile_resolved = True
             diagnostics.llm_api_key_set = _api_key_set(llm)
-        except (FileNotFoundError, ValueError):
+        except FileNotFoundError:
             diagnostics.errors.append(
                 f"LLM profile {profile.llm_profile_ref!r} not found"
             )
+        except Exception as e:
+            # Keep the dry-run total: the store can raise filelock.TimeoutError
+            # (lock contention), OSError, or a validation error before its own
+            # handler runs. Surface those as a diagnostic instead of crashing
+            # the editor preview (#3719) — distinct from a definitively-missing
+            # profile above.
+            diagnostics.errors.append(
+                f"Could not load LLM profile {profile.llm_profile_ref!r}: {e}"
+            )
     else:
         (
             diagnostics.acp_api_key_secret_name,
@@ -350,7 +359,12 @@ def resolve_agent_profile_dry_run(
         # diagnostics rather than raising, matching the API contract.
         try:
             if isinstance(profile, OpenHandsAgentProfile):
-                assert llm is not None
+                # valid here implies the LLM load above succeeded; gate
+                # explicitly rather than via assert (stripped under python -O).
+                if llm is None:
+                    raise RuntimeError(
+                        "OpenHands profile marked valid without a resolved LLM"
+                    )
                 settings = _build_openhands_settings(profile, llm, filtered_mcp, cipher)
             else:
                 settings = _build_acp_settings(profile, filtered_mcp)

tests/sdk/profiles/test_resolver.py

@@ -345,6 +345,29 @@ def test_dry_run_reports_dangling_llm_and_mcp(
     assert diag.resolved_settings is None
 
 
+def test_dry_run_total_on_llm_store_transient_error(
+    llm_store: LLMProfileStore, mcp_config: MCPConfig
+) -> None:
+    # The store can raise filelock.TimeoutError (lock contention) before its
+    # own handler runs; the dry-run must surface that as a diagnostic, not
+    # crash the editor preview (#3719).
+    def _boom(*_args: object, **_kwargs: object) -> LLM:
+        raise TimeoutError("profile store lock acquisition timed out")
+
+    llm_store.load = _boom  # type: ignore[method-assign]
+    profile = OpenHandsAgentProfile(
+        name="oh", llm_profile_ref="default", mcp_server_refs=["fetch"]
+    )
+    diag = resolve_agent_profile_dry_run(
+        profile, llm_store=llm_store, mcp_config=mcp_config, cipher=None
+    )
+    assert diag.valid is False
+    assert diag.llm_profile_resolved is False
+    # Reported as "could not load" (transient), distinct from "not found".
+    assert any("Could not load LLM profile" in e for e in diag.errors)
+    assert diag.resolved_settings is None
+
+
 def test_dry_run_verdict_matches_real_resolve(
     llm_store: LLMProfileStore, mcp_config: MCPConfig
 ) -> None: