Fix Linux terminal fd leak: bump node-pty to close inherited fds

[srid]

Long-lived Kolu sessions on Linux leak a /dev/ptmx master fd into every shell spawned after the first, and those leaked masters ride along into each shell's children — the suspected cause of the frozen split-terminal after a raw-mode TUI exits (#1076), plus a steady fd leak that pressures EMFILE in busy sessions. The root cause was a stale node-pty fork: our pin (fix/darwin-pty-fd-leak) branched 2025-12-21 and sat 127 commits behind upstream, predating the Linux fix for exactly this.

This repoints node-pty to a refreshed fork branch and drops a fork patch that has since landed upstream.

What changed

	Before (fix/darwin-pty-fd-leak)	After (feat/foreground-pid-dist)
Base	upstream @ 2025-12-21	upstream main (2026-05-18)
Linux fd-close fix (52417d516)	❌ missing	✅ included
foregroundPid accessor	✅ (fork patch)	✅ (same patch, now also PR #918)
macOS /dev/ptmx fix (#882)	fork patch	dropped — now upstream
built lib/ for git install	✅	✅

Upstream 52417d516 closes inherited fds in the forkpty() child before exec (close_range(2) → /proc/self/fd fallback). macOS was never affected (POSIX_SPAWN_CLOEXEC_DEFAULT), which is why this bug is Linux-only.

Verification

Measured on this branch with the rebuilt addon:

• Leak gone: a 2nd shell spawned after the first now reports 0 leaked /dev/ptmx fds (it was 1 before — confirmed in a real Kolu split sub-terminal, whose shell held fd 29 -> /dev/ptmx, the main terminal's master).
• foregroundPid intact: resolves to the foreground pgid (== pid), so pty-host's spawn-time sanity check still passes — agent/Dock foreground detection unaffected.
• just check clean; native pty.node compiles under nix (node-gyp rebuild).

Fork housekeeping

The upstreaming of foregroundPid is microsoft/node-pty#918 (rebased here, mergeable, awaiting maintainer review). Once it merges, the fork shrinks to just the committed-lib/ packaging shim — and could be retired entirely if we build node-pty from source in the nix derivation.

Refs #1076.

Generated by Claude Code (model claude-opus-4-8).

[srid]

Evidence

The fix is a resource-leak fix, so the proof is a measurement, not a pixel diff.

fd-leak: before → after (reproducible)

Spawn 6 PTYs sequentially (as the kolu server does), then count each shell's leaked /dev/ptmx master fds via /proc/<pid>/fd:

shell	BEFORE — fix/darwin-pty-fd-leak (cc137d90)	AFTER — feat/foreground-pid-dist (999e2bd)
1	0	0
2	1	0
3	2	0
4	3	0
5	4	0
6	5	0
TOTAL	15	0

Each shell N inherits the master fds of all N−1 terminals opened before it — and every process that shell spawns (agents, TUIs, build tools) inherits them too. The leak is O(n²) in a long session. With the bump: zero.

Cross-checked against the running app earlier in this investigation: a split sub-terminal's shell held fd 29 -> /dev/ptmx (PTMX_COUNT=1) — the main terminal's leaked master — reproducing upstream #710 ("every new terminal opens one more fd to /dev/ptmx").

foregroundPid preserved

The fork's load-bearing accessor still resolves on the rebased branch:

foregroundPid: 4190634 | pid: 4190634 | typeof: number

so pty-host's spawn-time sanity check passes and agent/Dock foreground detection is unaffected.

CI

justci — all 22 nodes succeeded across both lanes (x86_64-linux on an ephemeral box, aarch64-darwin on the static host):

• ci::e2e ✅ both platforms (Linux: 396/396 scenarios, 3221/3221 steps)
• ci::nix ✅ both — node-pty native compiles in the nix sandbox
• ci::pnpm-hash-fresh ✅ both — the refreshed pnpmDeps hash is reproducible
• ci::unit, ci::smoke, ci::home-manager, ci::biome, ci::fmt, ci::surface-example-build ✅ both

Measurement harness: spawn N PTYs, readlink /proc/<shell-pid>/fd/*, count targets matching /dev/ptmx. Run on Linux (the affected platform; macOS uses POSIX_SPAWN_CLOEXEC_DEFAULT and never leaked).

Generated by Claude Code (model claude-opus-4-8).