boot: page-cache, the host's cache and the guest's, measured

[aledbf]

Adds a lab probe, task boot:page-cache (lab option boot:page-cache), that answers three questions nothing in the repo measured:

1. Cold boot vs warm boot. Before each cold boot, the release's files are dropped from the host's page cache one file at a time with POSIX_FADV_DONTNEED (never drop_caches, which would also drop the other runners' cache). The probe interleaves cold and warm boots, reports p50/p95, and how much of the base image a cold boot reads.
2. Double caching. The guest writes and reads a file. The probe measures how much of the overlay is in the host's page cache (mincore), with the default disk mode and with --disk-direct-over-backing.
3. Whether the guest's page cache comes back to the host. The probe records QEMU's RSS at each step: idle, file written, guest cache full, guest drops its cache, file removed.

The overlay is created at 2 GiB and the guest's root filesystem is grown into it, so the file has room; the base image is not touched. A failed write in the guest fails the probe instead of being hidden.

One run on a laptop (REPS=1, v20260930.02), for orientation only; the lab runner's numbers are the ones to read:

	writeback	direct overlay
overlay in host cache after a 512 MB write	514 MB, kept after the guest drops and removes the file	2 MB
write 512 MB + fsync	963 ms	350 ms
re-read with the guest's cache empty	52 ms	112 ms
QEMU RSS, guest cache full → guest drops it	741 → 231 MB	738 → 212 MB

A cold boot reads 47 MB of the 830 MB base image.

Checked here: task lint, task test, and the probe end to end.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}