Windows terminals report exit about 1 s late: shorten node-pty's 1000 ms flush wait in T3's patch

[Pyseph]

On Windows, typing exit in a T3 terminal prints the shell's output at once, but the terminal only reports exited about 1 s later. "Process exited" and the tab auto-close wait that second, as do the onboarding install terminal and the mobile terminal screen.

The wait comes from node-pty's Windows agent. ConPTY keeps its output pipe open until the pseudoconsole is closed, so after the shell exits node-pty waits a fixed FLUSH_DATA_INTERVAL for trailing output before it destroys the socket. That interval is 1000 ms. node-pty raised it from 20 ms in 2020 (e430d82e, part of the ConPTY close-deadlock fix for microsoft/node-pty#375) and hasn't revisited it since.

Proposal: two changes to lib/windowsPtyAgent.js in T3's existing patches/node-pty@1.2.0-beta.15.patch:

1. FLUSH_DATA_INTERVAL goes from 1000 to 100 ms.
2. When the timer fires, the cleanup runs one setImmediate turn later, and only if no data event re-armed the timer in between.

The second part is what makes the first safe. Node runs timers before the poll phase, so with only the constant lowered, a main-thread stall longer than the interval right after the exit callback lets the cleanup discard output that is already in the pipe. cmd's final line was lost in 20 of 20 runs once the stall reached the interval. The extra turn lets buffered data be read first, and each chunk re-arms the timer, so a stall costs only time.

Measured through the real terminal Manager and NodePtyAdapter on Windows 11 build 26200, Node 26.12.1, from typing exit to exited, n=10 alternating, median (min):

Shell	Today	Patched
PowerShell 5.1	1,034 ms (1,024)	138 ms (123)
cmd	1,016 ms (1,007)	108 ms (106)

• Output loss: no patched run lost output: final lines, 1-50 MB writes, a 300 ms pause before the last line, main-thread stalls of 30-1,100 ms injected right after the exit callback, and a CPU-starved conhost.
• Real tails: with the stock timer, the longest output after the exit callback was 20 ms for cmd and 8 ms for PowerShell, over 240 sessions each.
• Stock node-pty: it loses cmd's last line in 18 of 20 runs at a 1,100 ms stall, which the patch also fixes.
• Tests: a new unit test covers the timing, and 4 of its 5 cases fail on main. src/terminal passes, 105 tests including the new ones.
• Other platforms: Linux and macOS never load this file.

Why I'm asking first:

• It changes when exited fires on Windows.
• Output that conhost emits more than about 100 ms after the shell's exit callback would be dropped; today that limit is 1000 ms.
• Not tested: Windows before 24H2, ARM, PowerShell 7 and the packaged app.
• With a CPU-starved conhost, several PowerShell runs timed out in the harness before exit was sent, so those cells are inconclusive.

The real fix is upstream in node-pty: microsoft/node-pty#990 and #991, both open, close the pseudoconsole on natural exit. They need native changes and new prebuilds, which T3 can't take through patches/. If either lands, this patch should go.

Is a patch-level fix like this acceptable until then? If so, I'll open a focused PR from Pyseph:perf/windows-terminal-exit-flush: the patch file, its regenerated lockfile hash and the test.

Written by Claude Opus 5.5 in T3 Code (Claude Code harness).

[Pyseph]

The change is up as #30880, so it can be judged on the code. Comments here or on the PR both work, and if the direction is wrong I'll close it.