api_dispatch: retry 529 OverloadedError as transient + transport_loss split on judge paths (rule-24 sibling library fix)

Overview / Motivation

Filed from the #952 china-politics-topup inline round (2026-07-15): the judge leg's initial dispatch lost 144/144 draws (48 divergence + 96 refusal) to HTTP-529 Overloaded errors that FAILED FAST instead of retrying — the known rule-24 library gap (.claude/rules/llm-judging.md rule 24), previously observed in #1090 (~2,638 stored 529 rows). #1206 codified the RULE; this task is the deferred sibling LIBRARY fix the rule names.

Goal

Make 529 OverloadedError a retried transient in api_dispatch.py, and add the rule-24 transport-loss plumbing (per-row transport counter + batch-collection re-dispatch of transport-class failures).

Bug observed

src/explore_persona_space/llm/api_dispatch.py:580 catches _anthropic.InternalServerError with the comment "includes 529 OverloadedError" (also :570-571 docstring) — but in the installed SDK (anthropic 0.88.0) OverloadedError's MRO is OverloadedError → APIStatusError, and issubclass(OverloadedError, InternalServerError) is False, so 529s bypass the transient tuple and fail fast.

• verified-at-filing: grep -rn "OverloadedError" src/explore_persona_space/ → 2 hits, both in llm/api_dispatch.py (:570 stale docstring, :580 stale inline comment on the InternalServerError tuple entry); no OverloadedError entry in the transient tuple (2026-07-15). Empirical: #952 topup judge leg, 144/144 529s failed fast, recovered only by a manual rule-24(ii) re-dispatch.

Proposed change (refine in planning)

Per rule 24's named sibling fix, .claude/rules/llm-judging.md:

1. Add _anthropic.OverloadedError (or APIStatusError with status_code == 529) to the transient retry tuple in api_dispatch.py; fix the stale :570/:580 comments.
2. Transport-class re-dispatch on the batch collection path (eval/batch_judge.py currently persists error: True rows; eval/graded_judge.py::_score_from_parsed folds them into content drops).
3. A separate transport_loss counter in JudgeResult so per-arm drop reports can split content-drops vs transport-losses (rule 24(ii)/rule 18).

Scope / surfaces

• src/explore_persona_space/llm/api_dispatch.py (transient tuple + comments)
• src/explore_persona_space/eval/batch_judge.py, src/explore_persona_space/eval/graded_judge.py (transport-vs-content split + re-dispatch)
• Tests pinning: 529 retried; exhausted 529 surfaces as transport_loss, never a content drop; batch collection re-dispatches transport-class rows.

Constraints / invariants

• Keep the AIMD 429 handling and existing cache keys byte-compatible (rule 22/23 cache caveats: do not change rubric-level cache keys).
• Bounded retries (existing max_attempts); fail loud on HTTP 400 invalid_request (rule 24(iii) unchanged).

Provenance

Evidence: #1090 (rule 23/24 incident), #1206 (rule change), #952 china-politics-topup GPU-leg report 2026-07-15 (144 fast-failed 529s, 0 after manual re-dispatch). Rule text: .claude/rules/llm-judging.md rule 24 + its "Companion note (library, sibling task)".