feat(learning-harness): the learning subcommands

Task 3.9. Five entry points make the classified detector set reachable:
classify a corpus and report the classifier against itself, score a catalog by
family, run round zero, run one evolve round, and publish every number from its
artifacts. The pairs command gains a runs-per-pair flag, since the family scorer
needs more than one.

Every step degrades with a reason rather than failing: with no chat endpoint the
run says so and still records what it could measure without a detector, and a
missing configuration directory is written from the current hunter prompt rather
than being an error. Artifacts are keyed by model throughout, so comparing a
second detector is a flag rather than a rewrite.

Both matrices are green and all three gate outputs stay byte-identical.

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

src/openultrasast/learning/rounds.py

@@ -241,7 +241,7 @@ def run_learning_round(
     sweep_families: Sequence[str] = (),
     spent_usd: Callable[[], float] | None = None,
     trajectories: Callable[[], Sequence[Mapping[str, object]]] | None = None,
-) -> object:
+) -> RoundRecord:
     """One round: propose one change for one family, then let the evidence decide whether it stays.
 
     Staged so that the cheap answer comes first: a train minibatch, then the family's held-out pairs,
@@ -437,7 +437,7 @@ def _finish(
     scores: Sequence[PairFamilyScore] = (),
     version: str | None = None,
     trajectories: Callable[[], Sequence[Mapping[str, object]]] | None = None,
-) -> object:
+) -> RoundRecord:
     """Write the round directory and journal the outcome. Every outcome is journalled, not just the good ones."""
     from ..redaction import redact_secrets