fix(learning-harness): the round-3 review findings, remediated RED-first

Eight of nine group-3 tasks were rejected. Each defect was hiding a number,
and the numbers are why the loop was not learning:

- The fixed side was judged against the *vulnerable* side's function
  ranges. 191 of 227 pairs with a resolvable labeled span (84%) have
  different spans on the two sides, so a leak was recorded as silence,
  always in the direction that flatters Youden.
- Round zero measured every family with the generalist. `family:unknown`
  earns no credit against any real family, so every family but `unknown`
  baselined at a structural zero and every noise floor with it.
- The cross-family sweep counted absolute failures instead of negative
  flips. On a corpus where most pairs fail at round zero that rejects
  every round for damage it did not do. Unscorable rows voted too.
- K runs was declared and never threaded; `pairs --k-runs` was parsed and
  discarded, and its only test grepped `--help`.
- The meta-agent proposer was a stub, so `learning round` could never
  propose anything. It now composes MetaAgent.evolve against a scratch
  copy of the family directory, and `--proposals FILE` runs a recorded
  sequence offline.
- `refused_holdout` was a vocabulary word no code path could produce, and
  the covering test accepted either outcome.
- Cost was metered once, after everything had been paid for; a stage that
  raised left no record at all.
- The corrected lever number was published beside the earlier one rather
  than in place of it, and the test named for that property never checked.
- Refusals were `continue` statements: `Refusal.degradation()` existed and
  nothing called it. A train-split identical twin was reported as
  `split:train`, which is untrue.

Test repairs that each changed a number: two split tests measured an
`x = 1` fixture that teaches nothing whatever the rule says; two CLI
scoreboard tests asserted against all-zero output produced by a stub that
answers twice and puts its finding outside the labeled function.

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

src/openultrasast/semantic/loo.py

@@ -99,11 +99,14 @@ def evaluate_loo(cases: Sequence[PairCase], *, facts: SemanticFacts | None = Non
             skipped.append((case.name, "pointer_pair_not_cached" if not case.vendored else "excerpt_missing"))
         else:
             targets.append(case)
-    from ..learning.split import is_teacher
+    from ..learning.split import is_teacher, refuse_if_holdout
 
     # learning-harness Req 5.1: every pair is still scored, but only a train-split teacher may seed a store.
     lessons = {case.name: pair_lessons(case, loaded) for case in targets if is_teacher(case)}
     outcomes = [_hold_out(case, targets, lessons, loaded) for case in targets]
+    # Req 5.2: the pairs that were kept out of the teaching set are named, not silently dropped.
+    refusal = refuse_if_holdout([case.name for case in targets if case.name not in lessons], targets)
+    refusals = (refusal.degradation(),) if refusal is not None else ()
     return LooResult(
         outcomes=tuple(outcomes),
         per_slice=_group(outcomes, lambda item: (item.slice,)),
@@ -112,7 +115,7 @@ def evaluate_loo(cases: Sequence[PairCase], *, facts: SemanticFacts | None = Non
         per_obligation_kind=_group(outcomes, lambda item: item.obligations),
         skipped=tuple(skipped),
         teaching_pairs=sum(1 for shapes in lessons.values() if shapes),
-        degradations=_unique_degradations(outcomes),
+        degradations=(*refusals, *_unique_degradations(outcomes)),
     )
 
 

src/openultrasast/semantic/seed.py

@@ -196,7 +196,9 @@ def _skip_reason(case: PairCase) -> str | None:
     from ..learning.split import is_teacher
 
     if not is_teacher(case) and case.review_tier in GATING_TIERS and case.vendored and not case.known_limit:
-        return f"split:{case.split}"  # learning-harness Req 5.1: only train-split, scorable pairs teach
+        # learning-harness Req 5.1: only train-split, scorable pairs teach. Say which of the two it was: two
+        # vibe-py rows are train-split identical twins, and calling them `split:train` is simply untrue.
+        return f"unscorable:{case.unscorable}" if case.unscorable else f"split:{case.split}"
     if case.review_tier not in GATING_TIERS:
         return f"tier:{case.review_tier}"
     if case.known_limit: