server: resolve Responses response ids to KV prefixes

[perfloop-agent]

Summary

• Add a bounded, server-local response-state index for /v1/responses. Compatible append-only requests with previous_response_id restore the exact KV frontier and parse, render, and tokenize only their new input tail.
• Generate the same id before streaming or buffered serialization, retain tool-call bindings and prompt configuration, and persist eligible frontiers at the existing continued-cache cadence so retained ids can resume after restart.
• Keep full replay for first requests, edits, configuration mismatches, and unavailable or evicted state. Response ids are opaque local handles rather than authorization; a missing retained checkpoint after restart asks the client to replay.

Test plan

• make -B cpu -j2
• make -B ds4_test -j2 && ./ds4_test --server
• make -B tests/responses_replay_bench -j2 && ./tests/responses_replay_bench --turns 171 --iterations 20 --disk-cache-mb 1024 --continued-interval-tokens 8192 --checkpoint-tokens 8198 --checkpoint-bytes 136806400 --expect-checkpoint-saved
• make -B tests/responses_replay_bench -j2 && ./tests/responses_replay_bench --turns 171 --iterations 2 --disk-cache-mb 1024 --continued-interval-tokens 8192 --checkpoint-tokens 8198 --checkpoint-bytes 136806400 --restart-check --expect-checkpoint-saved
• make -B tests/responses_replay_bench -j2 && ./tests/responses_replay_bench --turns 171 --iterations 2 --disk-cache-mb 64 --continued-interval-tokens 8192 --checkpoint-tokens 8198 --checkpoint-bytes 136806400 --expect-checkpoint-skipped

The benchmark uses the token-only test fixture, so it covers host-side request handling and continuation state rather than model or GPU inference.

Workload: in-memory steady-state 8k-token Responses response-id host ingress

Metric	Before	After	Change
host_ingress_ns/op	3164948	425542	86.6% lower
Checks: 5 passed.

---

Generated by Perfloop. Human sponsor: Tomas Senart. Measurements and checks.