# Incident — machine freeze, 2026-08-15 13:46 (Jetsam)

The third stacked-load freeze of this MacBook Air (after 2026-06-03 and
2026-07-10), and the first with a complete diagnostic chain on disk. The rules
in CLAUDE.md § "Machine limits" are derived from this file.

## What happened, with evidence

- **A `next dev` server outlived three sessions.** The member web app's dev
  server was started in session s30 (**Aug 13**) and left running through s32
  and s33 — roughly two days — while the Eve-polish harness drove hundreds of
  Playwright page loads and recompiles through it (phases 12.5–12.7 alone
  navigated 20 admin routes × repeats, 10-context probes, and three full lock
  suites). Its log (`/tmp/claude-eve-s30-web.log`) shows it still serving at
  **13:44:26 on Aug 15** — two minutes before the freeze.
- **`next dev` leaks, and its workers died of it.** Three node crash reports
  (`~/Library/Logs/DiagnosticReports/node-2026-08-15-0850*.ips`, `085401*.ips`)
  show fresh worker processes launched 08:49–08:50 aborting within 40s–3.5min at
  `v8::internal::V8::FatalProcessOutOfMemory` — JavaScript-heap OOM — with
  4.8–5.9 GB writable regions. That is the crash-respawn-recompile loop: each
  respawn recompiles the app, which is itself a memory and CPU spike.
- **The disk confirmed the thrash.** An "excessive disk writes" diagnostic fired
  at 09:06 (`disk writes_2026-08-15-090652`), and by 13:44 the dev server's
  filesystem-cache writes were taking **11.8–13.6 s** each.
- **Jetsam ended it.** `JetsamEvent-2026-08-15-134604.ips`: with 743 processes
  tracked, ~2 GB swap effectively full (1.15 GB still resident at idle after
  recovery) and the compressor at 4.47 **billion** compressions, macOS
  jettisoned a wave of system daemons — `trustd`, `secd`, `accountsd`, `cloudd`,
  `UIKitSystem`, … Killing those is what a full UI freeze looks like. Uptime
  shows **no reboot** (24 days): this was a memory-pressure freeze, not a kernel
  panic.

## The early warning that was visible and ignored

Near the end of s33, `grep` invocations against files that demonstrably
contained the pattern started returning **empty output** (a Python read of the
same file found the matches). A machine under memory pressure fails weirdly
before it fails loudly. That is a stop-the-line signal, not a curiosity.

## Why the rules already on file did not prevent it

CLAUDE.md already said "tear down at task boundaries, not session end" and
"production over dev for long Playwright runs". Two failure modes got around
them:

1. **The harness recipe outranked the hygiene rule in practice.** The Eve-polish
   stack recipe (memory + TODOS) pins dev servers for hot-iteration, so each
   session followed the recipe and inherited the long-lived server instead of
   restarting it.
2. **"Task boundary" was read as "phase boundary I felt like honoring".** The
   admin server was torn down at its boundary; the web server and BFF were not,
   because they were "still needed soon".

The rewritten rules therefore bind to observable events (session end, phase
close, N page loads, a memory reading) rather than to judgment.
