Instrument TCP write backpressure

[noise64]

• resolves https://linear.app/golem-cloud/issue/GOL-360
• resolves https://linear.app/golem-cloud/issue/GOL-389
• adds private P2/P3 profiling for JavaScript write batching, native socket crossings, copied/completed bytes, capacity waits, and wasi:http outgoing bodies
• keeps Rust ownership of pending write bytes and exposes the remaining count privately through write_queue_size()
• rearms JavaScript socket timeouts only after a strictly decreasing pending-byte sample; unchanged or increasing progress emits timeout
• uses native P2 timeout checkpoints because capacity waits can block JavaScript timer dispatch, and clears progress state on cancellation, close, and terminal paths
• retains the measured GOL-360 result: large-buffer delivery takes roughly 54–63 seconds and the pipe workload roughly 91–97 seconds even though its 5 MiB wasi:http body write completes in milliseconds
• leaves throughput ownership unresolved pending host-transport investigation; the timeout correction does not claim wire acknowledgement

Node compatibility

• primary CI-enforced remains 3174/4387 (72.4%)
• full public remains 3174/5750 (55.2%)
• slow buffer and pipe cases remain measured known gaps rather than receiving inflated per-entry timeouts

Focused P2/P3 timeout, pipelining, writev, formatting, and inherited stream-lifecycle checks pass at exact SHA 45f53e46bb30913711f18c413ba15850182815cb.

[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.