JS-104: Socket.write FIFO grows linearly — O(n^2) memcpy under backpressure

Proposed by the 2026-07-07 jab multi-agent review; confirmed by both skeptic verifiers. Buf.grow allocates exactly the requested size, and the socket write path requests size+len+64K, so a slow peer makes roughly every other write recopy the whole FIFO.

Input

Context

JABC_NET_JS Socket.write: this._wb.grow(this._wb.size + u.length + (64 << 10)) (jab net.cpp:285); Buf.grow (buf.cpp:100-108) allocates the exact size and copies all live bytes — no geometric doubling.
Buffering 100 MB in 64 KB writes to a stalled peer does ~800 grows copying ~40 GB total, one whole-buffer GC allocation each.
The 256K backpressure return (net.cpp:289) is only a Node-style hint; no caller in the repo checks it and zero 'drain' listeners exist.

Goals

Geometric growth: grow(Math.max(cap * 2, size + u.length)) (in the bundle or in Buf.grow itself); a microbench or test showing linear total copy cost.

Constraints

Keep the Node write()/drain contract; Buf.grow is shared — check other Buf users before changing its policy globally.

WIP

TODOs

decide: fix in Socket.write only, or make Buf.grow geometric for all
measure before/after with a stalled-reader TCP test