fix(benchmarks): robust retry/backoff for transient provider 503s

[Davincc77]

Summary

The first real v4.1 Gemini pilot produced 4 transient 503 UNAVAILABLE high demand failures. Existing retry was retry_max=2 with a flat ~1s backoff and no error classification, so transient infrastructure flakes surfaced as noisy benchmark failures.

This PR adds a small, scientifically controlled retry layer before any larger pilot/full runs:

• Classifier (providers.base.is_transient_error + TransientProviderError) recognises HTTP 429/500/502/503/504, gRPC UNAVAILABLE / RESOURCE_EXHAUSTED / DEADLINE_EXCEEDED, read timeouts, and overloaded/high-demand signatures.
• Gemini adapter re-raises classified SDK exceptions as TransientProviderError so the executor only retries on conditions where a retry can plausibly succeed.
• Executor uses exponential backoff with jitter, capped at retry_backoff_max_s. Permanent errors (auth, config, schema) abort immediately and are still recorded — retries do not hide them.
• Structured logs now include retried_attempts, per-attempt error_class/error_type/sleep_s, final_error_class, and cumulative_retry_delay_s in both raw_outputs.jsonl and errors.jsonl. The run manifest records the full retry configuration.
• CLI exposes --retry-max (default 5, cap 8), --retry-backoff (default 2s, cap 10s), --retry-backoff-max (default 30s, cap 30s), --retry-jitter (default 0.25).
• Workflow exposes retry_max, retry_backoff, retry_backoff_max workflow_dispatch inputs, validates and caps them, and surfaces them in the job summary. Low default concurrency is unchanged.
• Docs: README and BENCHMARK_PROTOCOL document the new retry/backoff defaults and the transient-vs-permanent contract.

Concurrency defaults stay low (default 1, workflow cap 2); the retry layer prefers waiting over hammering.

Testing

• python3 -m pytest benchmarks/v4.1/tests/ -q → 66 passed (18 new in test_retry_backoff.py).
• New tests cover: classifier on 503/429/RESOURCE_EXHAUSTED/timeouts and rejection of auth/config strings; transient-then-success path; persistent-failure surfacing with full trace; no-retry on permanent auth/config errors; backoff cap; jitter bounds; runner caps on --retry-max / --retry-backoff-max; manifest records retry settings.
• No real LLM calls — all tests inject scripted mock providers.
• yaml.safe_load validates the updated workflow.

No publish / no tag / no release / no Zenodo / no npm / no PyPI.

---

🤖 Generated by Computer