fix(cli): three bug-bash findings — live-socket detection, temp leaks, NaN/Inf

[WaylandYang]

Bug-bash session against last week's cleanup + Snapshot Hub work. Found three real issues; this PR fixes all three with retests on the dev box.

1. `forkd cleanup` would nuke live VMs (CRITICAL safety)

`workdir_has_live_process()` used `lsof ` to detect in-use work_dirs. On Ubuntu 24.04 / lsof 4.95, `lsof` against a Firecracker UNIX domain socket emits warnings to stderr and zero rows on stdout, even while a process actively holds it. Our code redirected stderr to /dev/null and trusted empty stdout to mean "no one is using this" — would have nuked a live VM's socket directory under `forkd cleanup --yes`.

Fix: replaced with a `/proc/*/cmdline` scan. Firecracker children pass `--api-sock /tmp/forkd-fork-/child-N.sock` on argv, so the work_dir path appears verbatim in cmdline while the VM is alive. Errs on the side of "live" if /proc is unreadable.

Reverified: the previously-misclassified `forkd-fork-pwb` (with two live firecracker children) now correctly shows `SKIP ... (live socket — a forkd run looks active)`.

2. `forkd unpack` leaked `/tmp/forkd-unpack-/` on failure

On any error after `create_dir_all(&tmp)` (corrupted tar.zst, truncated archive, sha256 mismatch, dest-exists-no-force), the temp extraction dir was never removed. Two such dirs were already sitting on the dev box, 16 MiB combined.

Fix: refactored into `unpack_into()` so the cleanup is a single `if result.is_err() { rm tmp }` wrapper. Same pattern applied to `pull_cmd` which had the parallel leak for the downloaded `.tar.zst`.

3. `forkd cleanup` didn't sweep `forkd-unpack-` / `forkd-pull-`

Only looked at `forkd-fork-` and `forkd-parent-`. Added the other two prefixes to a shared `PREFIXES` table; the "starts with /tmp/" + "name matches a known prefix" safety check used right before each remove uses the same table.

Bonus: warmup.js `Infinity` / `NaN` sentinels

`JSON.stringify(Infinity)` is `null` per JSON spec. Same for `NaN` / `-Infinity`. So `sb.eval("return 1/0")` came back as plain `null`, indistinguishable from a legitimately null return. Replacer now emits sentinel strings:

```
Infinity → "js_Infinity"
-Infinity → "js-Infinity_"
NaN → "js_NaN"
```

Takes effect on next rootfs rebuild of the recipe.

Verified on dev box (still-live pwb fork pair)

```
$ forkd cleanup
5 candidate work_dir(s):
SKIP /tmp/forkd-fork-pwb 4 KiB (live socket — a forkd run looks active)
DEL /tmp/forkd-fork-pyagent 3 KiB
DEL /tmp/forkd-unpack-2991226 0 KiB
DEL /tmp/forkd-unpack-2991309 16.0 MiB
DEL /tmp/forkd-unpack-test 0 KiB
dry run — pass `--yes` to delete the 4 dir(s) marked DEL above.

$ echo bogus > /tmp/bogus.tar.zst && forkd unpack /tmp/bogus.tar.zst
Error: tar entry
$ ls -d /tmp/forkd-unpack-* # no NEW dir added — old leaks plus pre-existing
$ forkd cleanup --yes
removed /tmp/forkd-unpack-2991226 / 2991309 / test

live fork untouched throughout

$ sudo -E forkd ping --child forkd-child-1
{ "pong": true, "agent_lang": "node", "warmup_ready": true }
```

Test plan

• cargo fmt + clippy + test (all pass)
• /proc scan correctly flags live firecracker work_dir
• unpack failure no longer leaks
• cleanup sweeps all four prefixes
• live VM survives `forkd cleanup --yes`

🤖 Generated with Claude Code