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

src/openultrasast/cli.py

@@ -1369,7 +1369,7 @@ def _learning_run(args: argparse.Namespace, cases: Sequence[PairCase], taxonomy:
         family=args.family,
         taxonomy=taxonomy,
         configs_dir=configs_dir,
-        proposer=_proposer(args, configs_dir, model),
+        proposer=_proposer(args, configs_dir, model, config.harnessx.provider),
         scan_factory=factory,
         model=model,
         journal=journal,
@@ -1402,13 +1402,13 @@ def _label(case: PairCase, taxonomy: object) -> str:
     return _family_of(case, taxonomy)  # type: ignore[arg-type]
 
 
-def _proposer(args: argparse.Namespace, configs_dir: Path, model: str) -> Proposer:
+def _proposer(args: argparse.Namespace, configs_dir: Path, model: str, provider: str = "anthropic") -> Proposer:
     """`--proposals FILE` runs a recorded sequence offline; without it the round asks the meta-agent."""
     from .learning.proposer import HarnessXProposer, Proposal, ScriptedProposer
 
     path = getattr(args, "proposals", None)
     if path is None:
-        return HarnessXProposer(configs_dir=configs_dir, model=model)
+        return HarnessXProposer(configs_dir=configs_dir, model=model, provider=provider)
     rows = json.loads(Path(path).read_text(encoding="utf-8"))
     return ScriptedProposer(
         [

src/openultrasast/learning/proposer.py

@@ -177,6 +177,9 @@ class HarnessXProposer:
 
     configs_dir: Path
     model: str
+    # The vendor the meta-agent talks to. Hard-coding anthropic here sent it to a provider with no key configured
+    # while the detector was talking to DeepSeek, and the round journalled that as "no proposal".
+    provider: str = "anthropic"
     reason: str = ""
     agent: MetaAgentCall | None = None  # injected in tests; None composes MetaAgent behind harness_ext
 
@@ -190,7 +193,7 @@ class HarnessXProposer:
                 self.reason = "harnessx_unavailable"
                 return None
             try:
-                call = _meta_agent_call()
+                call = _meta_agent_call(self.provider)
             except Exception:  # noqa: BLE001 — a proposer that cannot start proposes nothing and says so
                 self.reason = "harnessx_unavailable"
                 return None
@@ -263,7 +266,7 @@ def _model_config(model: str, provider: str = "anthropic") -> object:
     return ModelConfig(main=build_provider(model, provider))
 
 
-def _meta_agent_call() -> MetaAgentCall:
+def _meta_agent_call(provider: str = "anthropic") -> MetaAgentCall:
     """Compose `MetaAgent.evolve` behind `harness_ext`, writing only inside the scratch copy.
 
     `evolve` is a coroutine and returns a HarnessX config path; what this proposer takes from it is the state of the
@@ -279,7 +282,7 @@ def _meta_agent_call() -> MetaAgentCall:
         trajectories = workspace.parent / "trajectories"
         trajectories.mkdir(parents=True, exist_ok=True)
         (trajectories / "facts.json").write_text(facts_prompt(facts), encoding="utf-8")
-        agent = _meta_agent_class()(inner_model=_model_config(model), allowed_write_roots=(workspace,))
+        agent = _meta_agent_class()(inner_model=_model_config(model, provider), allowed_write_roots=(workspace,))
         output = workspace.parent / "meta-out"
         output.mkdir(parents=True, exist_ok=True)
         asyncio.run(agent.evolve(current_config=workspace, trajectories_dir=trajectories, output_dir=output))

src/openultrasast/learning/rounds.py

@@ -418,7 +418,18 @@ def run_learning_round(
     proposal = proposer.propose(facts)  # type: ignore[attr-defined]
     record = _round_record(round_number, family, proposal, taxonomy, config)
     if proposal is None:
-        return _finish(journal, directory, record, outcome="rejected", reason="no_proposal", cost=spent, trajectories=trajectories)
+        # `no_proposal` covers two different worlds: a proposer that considered the facts and declined, and one
+        # that could not start at all. Losing that distinction hides a fixable configuration error as a quiet round.
+        why = str(getattr(proposer, "reason", "") or "")
+        return _finish(
+            journal,
+            directory,
+            record,
+            outcome="rejected",
+            reason=f"no_proposal: {why}" if why else "no_proposal",
+            cost=spent,
+            trajectories=trajectories,
+        )
     refusal = refuse_proposal(proposal, config)
     if refusal is not None:
         return _finish(