Runtime: transport-aware retry policy for model connection failures

[jafreck]

Problem

• Migration runs repeatedly fail with transient upstream transport failures (HTTP/2 GOAWAY, CAPIError 503), causing task churn and manual reruns.

Proposal

• Add error classification for transient transport failures vs deterministic failures.
• For transient errors, apply exponential backoff with jitter and a per-model/per-agent circuit breaker.
• Keep deterministic failures on fast-fail paths.
• Emit structured retry telemetry (error_class, attempt, backoff_ms, breaker_state).

Acceptance Criteria

• Transient transport failures are retried with exponential backoff + jitter.
• Deterministic/schema failures are not endlessly retried.
• Circuit breaker opens after configurable threshold and cools down.
• Logs clearly show classification and retry decisions.
• Add/update tests for retry classification and breaker transitions.

Context reference

• Observed repeated GOAWAY/503 agent failures during zstd migration runs.