Make response caching safe for Anthropic streaming compaction blocks

[afourniernv]

Scope

This only affects Relay's optional response cache when an Anthropic request enables server-side compaction or carries a prior compaction block. Ordinary Anthropic traffic is out of scope, and the first live SSE stream is forwarded correctly.

The data-loss edge case occurs when a streamed request actually triggers compaction: cached replay cannot currently reproduce the native compaction_delta lifecycle. Related cache keys also omit the Anthropic beta/version headers, and replay must preserve missing versus explicit-null fields.

#1035 fixes the separate stream-aggregation bug.

Resolution

Until native replay and header-aware keys are implemented, bypass response caching for compaction-enabled and continuation requests. Partial, malformed, or legacy compaction entries must always miss and run live.

Anthropic compaction documentation

[afourniernv]

Narrow collector fix: #1035. This issue remains open for the separate cache-key migration, omission-preserving replay, and safe bypass removal.

[mnajafian-nv]

I’m taking the conservative cache-safety fix: bypass the optional response cache for Anthropic compaction-enabled requests and compaction continuations. Native replay and header-aware cache keys remain out of scope for this change.

[mnajafian-nv]

PR #1127 merged the conservative cache-safety fix. Requests that enable Anthropic compaction or continue from a compaction block now bypass the optional response cache and run live. This issue remains open for header-aware cache keys, faithful compaction replay, and eventually removing the bypass.