Instrument TCP write backpressure

[noise64]

Summary

• add private, feature-gated P2/P3 profiling for JavaScript write/writev batching, drain/completion events, native socket crossings, copied/completed bytes, P2 capacity waits, P3 write-all work, and wasi:http outgoing bodies
• add a direct/corked runtime matrix spanning 4 KiB, 64 KiB, 64 KiB + 1, and 1 MiB writes while preserving ordinary-feature runtime coverage in a distinct component artifact
• record the measured ownership boundary without selecting a speculative production mitigation

Measurements

• the large-buffer marker completes in roughly 54–63 seconds, with host-facing socket delivery dominating both previews
• the pipe marker takes roughly 91–97 seconds while its 5 MiB wasi:http body write completes in milliseconds, placing most wall time outside that transport path
• increasing the receiver request size reduced crossings but regressed P2 wall time, so that experiment was reverted

Validation

• exact-SHA CI passed all 37 executed checks, including all P2/P3 runtime groups and node-compat shards
• independent Codex and Claude delta reviews found no high- or medium-severity regression
• the slow buffer and pipe vendored cases remain measured known gaps because longer per-entry timeouts would add excessive concurrent-CI wall time
• node-compat counts remain unchanged: primary CI-enforced 3174/4387 (72.4%); full public 3174/5750 (55.2%)

Boundary and next action

This PR is the reviewable instrumentation/research tranche for GOL-360; it does not claim the production throughput fix. The measurements require a truthful host-facing progress/backpressure contract before GOL-389 timeout semantics can distinguish an advancing native write from a stall. No batching, timeout, retry, or production scheduling change is included.

[noise64]

Cycle-2 candidate 6cf399ff9b88ed4bf7c2e12e62562f20e9402f93 is now pushed.

• Codex delta review: clear; shared-target collision and timeout-deferral evidence resolved.
• Claude Opus delta confirmation: clear; no high/medium regressions.
• P2 runtime group 8: 54/54 passed, including the profiling matrix and ordinary HTTP lifecycle cases.
• P3 runtime group 8: 54/54 passed with the same coverage.
• Both lanes confirmed distinct ordinary and profiling component artifacts, closing the prior collision.
• Formatting, DTS currentness, node-compat report currentness, ordinary P2/P3 guards, profiling P2/P3 guards, and introduced-lint isolation pass.

The earlier CI failures belong to 0cec3780 and were caused by feature-off registration of the private bridge method; the pushed candidate fixes that compile failure.

Uninstrumented controls were P2/P3 large Buffer 55.821s/55.566s and pipe 91.377s/91.148s. The P3 pipe profile completed the 5 MiB wasi:http body in one write_all in 12.255ms in one run and 6.629ms in the final diagnostic run, while total fixture time remained 92-97s. The slow Buffer and pipe cases therefore remain measured known gaps based on concurrent-CI cost and reliability, not a fixed timeout limitation. No production batching, retry, timeout, or scheduling mitigation is claimed by this draft.

[noise64]

Exact-SHA readiness gates are complete for 6cf399ff9b88ed4bf7c2e12e62562f20e9402f93.

• all 37 executed CI checks passed, including every P2/P3 runtime group and node-compat shard;
• P2/P3 runtime group 8 passed in 11m21s/11m16s;
• the post-push Claude delta confirmation completed with verdict clear and no new high/medium regressions;
• release/publish and [code]smith were expected skips.

The draft is technically ready for maintainer review; no CI reruns or readiness-state mutation were performed.