fix(anthropic): reject truncated judge replies when requested

[rudycelekli]

Problem

The analytic judge explicitly sets ProviderConfig.reject_truncated=True, but Anthropic ignores stop_reason=max_tokens. Even a syntactically complete JSON prefix can be treated as a valid verdict despite the service declaring generation cut off. Empty cutoffs also become empty-content errors rather than the recoverable shared truncation signal.

Change

Use the existing shared truncation guard before empty-text checks, only when requested. Preserve partial SUT content by default and all complete judge replies.

Verification

• Real Anthropic SDK0.107.1 over owned HTTP transport: baseline3failed/3passed; patched6passed, zero skipped. Includes the actual analytic judge + evaluation pipeline with a parseable cut-off verdict and complete-verdict control.
• Worked production run_selected({B13}) with existing Bridge constant SUT transport and a local Anthropic-format server: baseline B13 PASS100%; patched B13 INCONCLUSIVE after one rejected judge generation. This deliberately synthetic run does not measure any live vendor/model, and its selected-suite overall grade remains withheld.

Illustrative owned-transport B13 scorecard
Before: B13 pass, score1.0; no judge transport failures
After: B13 inconclusive, score0.0 (unscored); owned-judge transport failures1
Reason: ProviderTruncatedError, finish_reason=max_tokens

Ruff/Bandit/layout/all11example gates run separately; unchanged fork CI validates those gates for Python3.10/3.11/3.12, not pytest. Optional SDK tests run locally. Advisory Mypy reports 23 diagnostics in three unchanged provider/resolver files and none in the changed adapter; a full typecheck pass is not claimed. AI assistance used.

Exact-head CI

All three unchanged fork CI jobs (Python 3.10/3.11/3.12) passed at signed+DCO head 11fa1d64c2b4403bf6f8f745ff69518398e1eb1d: workflow evidence. These jobs run installation, lint, Bandit, layout and fixture validation; the native pytest results above were run locally. Upstream workflow approval and maintainer review are separate.

[n-papaioannou]

Thanks Rudy, this works and matches what the HTTP and OpenRouter judges already do. We're holding it for a fix on our side first: right now one cut-off judge reply retires that judge for the rest of the run, so with no fallback every later probe goes unscored. Once that's fixed on main we'll merge this.

[rudycelekli]

Understood. Keeping this PR on the requested hold until main's judge recovery handles a cut-off reply without retiring the only judge for the remainder of the run. Current main e08f1ca8f481dff49f15f89c6df600920d438a2f still retires a model after JudgeTransportExhaustedError; no changes here override that decision. The provider cutoff patch can be revalidated after the upstream recovery fix.