perf(dataplane): steady-state fast path at the response point (v5 S2c, default-off)

[bb7133]

Stacks conceptually on S2b (command point, merged 2ec605dc). Extends the inline fast path to the response-ACK point, default-off behind TIPROXY_FSM_FASTPATH.

Why

Measured on current main: for the online write-heavy / small-single-packet-response model, each command has ~1 non-opt#5-bypassed await_effect(ForwardResponseToClient) round-trip — the same magnitude as the command-forward round-trip that S2b fast-pathed for −6.9% CPU. The response packet is already flushed to the client (session_engine.rs:3646) before this round-trip, so it is pure FSM bookkeeping with no I/O.

Mechanism (design converged with CodexM5)

The command's own forward permit is held across response_rounds (pending >= 1), so S2b's pending == 0 gate is unreachable. New core API try_inline_commit_excluding(&fwd_permit, event, expected, shutdown):

• fwd_permit is a capability/liveness witness — Rust's borrow proves it is not dropped; Arc::ptr_eq proves it belongs to this core, so pending == 1 means that single registration is exactly this permit (no redirect/drain/control/in-flight sync).
• Identity check, open/shutdown read, speculative clone transition, exact-effect match, write-back — all in one lock critical section (no raw pending escapes → no TOCTOU).
• Registers no new permit (no post-commit I/O); fwd_permit keeps the floor.
• Same-core mismatch / pending != 1 / sealed / shutdown / effect mismatch → unchanged slow path. LocalInfile always slow. Prepared/cursor commands keep an in-flight sync (pending >= 2) → auto-fallback.

Verification

• Flag-off keeps the exact slow-path ack-permit lifetime; fast-path on/off wire traces byte-identical (scripts/fastpath-trace-ab.sh).
• Response point fires once per single-packet-response command under the flag (0 off).
• New gate tests: inline commit (no new permit, same FSM state as slow transition, terminal round) + every fallback (pending != 1, wrong-gate witness, sealed, shutdown, effect mismatch).
• All dataplane suites green, fmt + clippy clean.

Hit-rate / CPU A/B on a production-like workload is a separate measurement step (as #146 was for S2b).

🤖 Generated with Claude Code