[eval] Exclude unresolved endpoint failures from benchmark scores

[penfever]

PR #51 marks exhausted endpoint requests with
[EVALCHEMY_INFRASTRUCTURE_ERROR], but the marker can still reach a benchmark
scorer as model text. Supersedes #50.

Done when:

• Retryable endpoint failures are retried using the same outcome policy as
  Harbor.
• An unresolvable failure is retained in samples and run output as
  infrastructure state, without contributing a benchmark miss or metric value.
• Regression coverage distinguishes retryable, unresolvable, and genuine empty
  model completions.

[penfever]

🤖 Related: #19 contains the Harbor-derived retry and error-classification proposal for the same robust_api.py path. Its proposed terminal policy scores exhausted transient, rate-limit, and unknown failures as misses; this issue requires those unresolvable failures to be retained as infrastructure state and excluded from benchmark metrics. Reconcile that policy before implementation.