minio/docs/investigations/r8/review/implementation-prompt.md

Read-only implementation review of SILO R8 on baseline 9ebe81c1b3. You previously agreed plan v1 SHA256 7cb609e37e3199ecd93c992f08d123968ba8082682a108f20490e0644d735366; this is a fresh actual Opus 5.0 review. Read docs/investigations/r8/plan-v1.md, consensus.md, review/opus-v1.md, review/implementation.patch (production diff SHA256 88716d9af433b55527308e09949b0adaab6980aca3bf1ad38b677e9c1c0ce083), and every current file in review/implementation-files.json. These are actual edits, including new untracked tests not represented in git diff. Do not edit or try to write a plan file; deliver findings in the response. Allowed tools are Read/Grep/Glob only.

Independent checks: exact phase ordering in current Go1.27.1 net/http, mutex/atomic interaction, no header cap renewal across >250ms updates, nonzero/zero/past handling, H1 body rolling, buffered/pipelined and keep-alive transitions, TLS/h2 and grid/hijack default-caller compatibility. Detect concrete bugs and weak tests, avoid hypothetical unrelated expansions. Also see evidence/implementation-focused.log, evidence/default-30s-transfers.log (>33s plaintext and TLS H1 uploads/downloads passed), evidence/grid.log, and evidence/linux-race.log. The first Darwin race run failed only because the H2 fixture's equal read/write timers raced to supply different timeout error types (not a data race); final fixture clears the per-stream write timer to isolate native ReadTimeout and checks healthy multiplexed requests and connection reuse. Final full race reruns are in progress; do not treat pending checks as passed.

Give Chinese verdict APPROVE or REQUEST_CHANGES, concrete file/line findings separated into blocking and nonblocking, and explicit implementation conformance to accepted v1. Clearly distinguish independently inspected source from tests you did not personally execute. If no blockers, say so.