perf: v1.1.0-beta.9 — 3.3M H2 rps, context leak fix, zero-alloc frame fast path

[FumingPower3925]

Summary

Profile-driven optimization loop targeting HTTP/2 throughput and io_uring parity with epoll. Two performance commits plus one infrastructure commit.

Commit 1: Fix H2 context leak + 9 optimizations (+190% io_uring H2)

• Fix H2 context pool leak: acquireContext cached contexts on ephemeral H2 streams, causing them to leak (never returned to sync.Pool). This was 87% of all H2 allocations. Fix: only cache on H1 streams (persistent keep-alive connections). H2 inline handlers use InlineCachedCtx instead.
• io_uring: reorder H2 drain before dirty list (reduces pipeline stalls)
• Async H2 adapter: use manual HPACK content-length encoding (~140ns/req)
• H1: fused cachedStatus200Date block (one append for status + date)
• io_uring: conditional immediate Submit (skip when CQEs already visible)
• Pre-allocate Stream.Headers capacity for H2 (avoid first-use allocation)
• Persistent HPACK emit function per Processor (eliminate per-request closure)
• Reorder IsCancelled after canRunInline (avoid atomic on hot inline path)
• Remove redundant nil check in acquireContext

Commit 2: Zero-alloc HEADERS fast path (+12% io_uring H2)

Bypass x/net framer for common HEADERS frames (END_HEADERS set, no PADDED/PRIORITY, not during CONTINUATION). The framer's *HeadersFrame allocation was 98% of remaining H2 allocations (2.5 GB/20s at 3M rps). The fast path reads the 9-byte frame header directly, extracts streamID/flags/payload, and passes them to ProcessRawHeaders — all RFC 7540 validations preserved.

Infrastructure: Three-way benchmark comparison

CloudBenchmarkSplit now builds 3 binaries (main, HEAD savepoint, current working tree) and runs a 9-pass interleaved schedule for three-way comparison. Enables tracking both total improvement (vs main) and incremental improvement (vs last commit).

Cloud Benchmark Results (arm64 c7g.2xlarge, split server/client)

Config	Before beta.9	After beta.9	Delta
io_uring H2	930K rps	3.30M rps	+255%
epoll H2	2.37M rps	3.33M rps	+41%
io_uring vs epoll H2 gap	-56 to -69%	±1%	Eliminated
H2/H1 ratio	160-400%	570-577%	H2 is 5.7x H1
H1 (all engines)	~580K rps	~590K rps	Stable

Profile Analysis (post-optimization)

• H1: 81% kernel syscall-bound, 0 allocations on hot path, fully optimized
• H2: 18% kernel + 27% HPACK decode + 27% handler/response + 13% sync.Pool + 9% stream mgmt
• H2 allocations: 18 KB total over 20s (down from 2.5 GB before zero-alloc fast path)

Test plan

• mage fullCompliance — all 4 phases pass (unit tests with race detector, 9 fuzz targets, h1spec + h2spec 142/146, conformance matrix 9 engine×protocol combos, integration tests)
• mage cloudBenchmarkSplit — 9-pass interleaved A/B/C on arm64 c7g.2xlarge
• mage cloudProfileSplit — CPU + allocation profiles on 4 configs
• h2spec: 146/146 on io_uring and epoll (4 known std engine failures unchanged)
• Zero H1 regression across all engine×objective combinations