fix(learning-harness): say why the proposer had nothing, and stop hard-coding its vendor

The meta-agent round journalled `no_proposal`. The real cause was an
authentication failure: `_model_config` hard-coded the anthropic provider
while the detector was talking to DeepSeek and no ANTHROPIC_API_KEY exists
here. The composition itself was fine -- MetaAgent constructed, evolve
started, budget $5 -- and the trace was thrown away.

`no_proposal` covered two different worlds: a proposer that considered the
facts and declined, and one that could not start at all. The journal now
carries the proposer's own reason, and the vendor comes from
`[harnessx] provider` instead of being assumed.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG

tests/test_learning_cli.py

@@ -161,7 +161,9 @@ def test_a_round_runs_from_the_command_line_and_journals_its_outcome(workspace:
     assert len(rounds) == 1 and rounds[0]["family"] == "injection"
     # Without the extra there is no meta-agent, so the honest outcome is exactly this one. Accepting any member of
     # the enum would have hidden that the round could never propose anything at all.
-    assert (rounds[0]["outcome"], rounds[0]["reason"]) == ("rejected", "no_proposal")
+    # The reason now carries *why* the proposer had nothing: "no proposal" and "could not start" are different
+    # worlds, and only one of them is a fixable configuration error.
+    assert rounds[0]["outcome"] == "rejected" and rounds[0]["reason"].startswith("no_proposal")
     assert (workspace / "learning" / "rounds" / "1" / "attribution.json").is_file()
 
 

tests/test_learning_publish.py

@@ -187,8 +187,13 @@ def test_two_slices_of_one_model_are_two_rows_not_one_overwriting_the_other(tmp_
                     "taxonomy_version": "1",
                     "metrics": {
                         "injection": {
-                            "family": "injection", "scorable": 10, "unscorable": {}, "recall": recall,
-                            "silence": 1.0, "youden": recall, "hierarchical_credit": False,
+                            "family": "injection",
+                            "scorable": 10,
+                            "unscorable": {},
+                            "recall": recall,
+                            "silence": 1.0,
+                            "youden": recall,
+                            "hierarchical_credit": False,
                             "outcomes": {"pair_correct": 5},
                         }
                     },

tests/test_learning_round4_evidence.py

@@ -68,12 +68,12 @@ def test_the_meta_agent_is_built_with_a_model_config_not_a_model_name(tmp_path:
             return workspace
 
     monkeypatch.setattr(proposer_module, "_meta_agent_class", lambda: _Agent)
-    monkeypatch.setattr(proposer_module, "_model_config", lambda model: f"ModelConfig({model})")
+    monkeypatch.setattr(proposer_module, "_model_config", lambda model, provider="anthropic": f"ModelConfig({model}@{provider})")
     (tmp_path / "injection").mkdir()
     (tmp_path / "injection" / "checklist.md").write_text("- old\n")
     proposal = proposer_module.HarnessXProposer(configs_dir=tmp_path, model="m").propose(_facts())
     assert proposal is not None and proposal.change == {"checklist.md": "- from the meta agent\n"}
-    assert seen["inner_model"] == "ModelConfig(m)", "the agent was handed a model name where it wanted a config"
+    assert seen["inner_model"] == "ModelConfig(m@anthropic)", "the agent was handed a model name where it wanted a config"
     assert seen["allowed_write_roots"] != (tmp_path / "injection",), "the live family directory is never a write root"
 
 

tests/test_learning_round_evidence.py

@@ -373,3 +373,36 @@ def test_round_zero_refuses_fewer_than_three_runs(tmp_path: Path) -> None:
             out_dir=tmp_path / "out",
             k_runs=1,
         )
+
+
+def test_a_round_journals_why_the_proposer_had_nothing_to_say(world) -> None:  # type: ignore[no-untyped-def]
+    """`no_proposal` covers two different worlds: a proposer with nothing to add, and one that could not start.
+
+    Measured: the meta-agent round journalled `no_proposal` while the real cause was an authentication failure in a
+    provider seam pointing at the wrong vendor. A round that failed for a fixable configuration reason must not
+    look identical to a model that considered the facts and declined."""
+
+    class Mute:
+        reason = "meta_agent_failed: RuntimeError"
+
+        def propose(self, facts):  # type: ignore[no-untyped-def]
+            del facts
+            return None
+
+    from openultrasast.learning.rounds import run_learning_round
+
+    record = run_learning_round(
+        world["cases"],
+        family="injection",
+        taxonomy=load_families(),
+        configs_dir=world["configs"],
+        proposer=Mute(),
+        scan_factory=_injection_improves,
+        model="m",
+        journal=world["journal"],
+        archive=world["archive"],
+        floors=world["floors"],
+        out_dir=world["out"],
+    )
+    assert record.outcome == "rejected"
+    assert record.reason == "no_proposal: meta_agent_failed: RuntimeError"