fix(interception): per-digest retry-replay cache

[faresobeid]

The single last_request/last_response slot assumes one outstanding model request per rollout session. A harness running concurrent conversation branches (e.g. rho's new parallel() subagent fan-out) breaks that assumption: completing any turn evicts every other branch's replay entry, so a marked SDK retry that should replay re-samples instead and forks the graph.

This keys the cache by request-body digest (OrderedDict, bounded at 32, FIFO eviction), preserving the exact replay-on-marked-retry and fresh-attempt-supersedes semantics per body while making them immune to interleaving. The streaming path is untouched (it never wrote the slots). tests/v1 green locally (e2e skipped, needs PRIME_API_KEY).

🤖 Generated with Claude Code

Note

Replace single-entry retry replay with per-digest cache in InterceptionServer

• Replaces RolloutSession.last_request and last_response with an OrderedDict cache bounded to REPLAY_CACHE_SIZE (32). Marked SDK retries now replay the recorded response for up to 32 distinct request bodies per session.
• A fresh attempt with a matching request body clears its prior replay entry.
• Risk: Removes RolloutSession.last_request and RolloutSession.last_response. Code accessing these must use RolloutSession.replays.

^{Macroscope summarized 3bd913d.}