Persist the #2151 api-refusal drop class in wave metas (currently blended into content_drops)

Goal

Make the api-refusal drop class (#2151) survive into the persisted wave meta as its own field, so a recoverable transport-conditional refusal is never indistinguishable from a legitimate content decline in the committed record.

The defect

src/explore_persona_space/eval/graded_judge.py computes the third top-level drop class (n_api_refusal_draws, per_item_api_refusals) and emits a precise WARNING naming it:

judge reduce: 13 API-refusal draws across 13 item(s) (stop_reason == 'refusal', empty
content) — the wave ran a transport whose safety classifier CENSORS; transport-conditional
and retriable via targeted SYNC re-issue at the identical instrument (llm-judging.md rule
28, #2151/#1739). NOT blended into content drops or transport losses.

But the PERSISTED <wave>.meta.json carries no api-refusal field. Measured on #2329 q35_ladder_decay Leg B pre-recovery metas — the only per-arm drop keys are truncation_drops and content_drops, and the only top-level drop key is residual_transport_lost:

dfrag-r1_pirate       grid-ceiling  n_items=480 n_scored=478 content_drops=2
dfrag-r2_butler       grid-ceiling  n_items=480 n_scored=479 content_drops=1
dfrag-r5a_lu_therapy  grid-ceiling  n_items=240 n_scored=239 content_drops=1
dfrag-r5b_lu_philosophy grid-steered n_items=200 n_scored=199 content_drops=1
dfrag-r5b_lu_philosophy grid-ceiling n_items=400 n_scored=388 content_drops=12

2+1+1+1+12 = 17, exactly the api-refusal count the WARNING reported. So the class is correct in the log and blended in the artifact.

Why it matters

The log line is ephemeral; the meta is the durable record every downstream consumer and every later reader uses. The two classes have OPPOSITE remedies:

• content drop — the judge legitimately declined on the content. Correct handling is to DROP the row and report it. Not retriable.
• api-classifier refusal — the provider's safety classifier censored a request that SUCCEEDED. Rule 28 says transport-conditional and retriable via targeted sync re-issue at the identical instrument; #1739 measured 0 re-refusals on sync, and #2329 Leg B reproduced that (17/17 recovered, zero re-refusals).

Blending them into content_drops makes a recoverable, transport-attributable loss look like an irreducible content property of the item. A future round reading these metas would conclude 17 items are simply unjudgeable and would NOT attempt the sanctioned recovery — which is the whole point of the #2151 class existing.

Aggravating factor: the wave-level resume predicate is <wave>.meta.json vs regime, so a wave with refusals reports complete (meta matches regime) — skip. The refusals are therefore both misclassified AND frozen behind a completion marker, with the only signal being a log line that scrolls away.

Proposed fix

Persist the class alongside the existing ones — per-arm api_refusal_drops and a top-level n_api_refusal_draws (plus per_item_api_refusals if cheap) in the wave meta written by the run_wave / graded-judge reduce path. Keep content_drops meaning content-only after the change.

Consider also recording the wave's realized TRANSPORT in the meta (batch vs sync): the refusal class is transport-conditional, so "which transport produced this row" is the other half of the provenance. #2329 Leg B ended with a deliberate, rule-28-sanctioned split (6,015 draws batch + 17 sync) that is currently recoverable only from prose.

Acceptance criteria

1. A wave whose reduce reports N api-refusal draws writes a meta in which those N are attributable to the api-refusal class and are NOT counted in content_drops.
2. Reproduce with the #2329 fixture: the committed pre-recovery metas at eval_results/issue_2329/q35_ladder_decay/decay/judge/prerecovery_quarantine/ (4 files, byte-identical copies) whose true api-refusal counts are 2 / 1 / 1 / 13.
3. A genuine content drop is still reported as content_drops — the fix must not reclassify content declines as refusals (the inverse error).
4. Back-compat: existing consumers reading content_drops must not silently change meaning without the field being explicit; state the migration posture for already committed metas (they under-report content_drops' true class — do NOT rewrite history).
5. Tests failing before / passing after; no new red in the no-flags workflow_lint.py run or the mapped-test selection.

Provenance

Found during #2329 round q35_ladder_decay Leg B (2026-08-21). The orchestrator's own epm:progress v199 note asserted "refusals correctly NOT blended into content drops or transport losses" — true of the log line, FALSE of the persisted artifact; the R-1 implementer caught the contradiction while building the drop-class diagnostics panel, and the orchestrator verified it directly against the quarantined metas before filing. That correction is recorded on #2329.

Evidence: #2329 events.jsonl v199 (the wrong claim) and the correction note; the committed pre-recovery metas named in criterion 2; graded_judge.py docstring lines ~173-185 (the class definition and precedence transport -> api-refusal -> content).

• target_file: src/explore_persona_space/eval/graded_judge.py
• fingerprint: api-refusal-class-not-persisted-in-wave-meta
• confidence: high — measured directly, counts reconcile exactly to the warned total