RLIMIT_FSIZE defaults to 1 GiB regardless of the disk size accepted in the same options, killing the VM with no crash record

[4ndr3wxh1ll]

Running at rev `8803834036205cf2cac5cfca98bb3875812c897a`.

What happens

A box created with an 8 GiB disk dies the instant its guest has written 1 GiB. The client sees:

Attach stream error: status: Unknown, message: "h2 protocol error: error reading a body from connection"

and the exec wait RPC reports `exit_code: -1`. Both are the client watching the whole microVM disappear, not a stream fault.

Root cause

The VM monitor is jailed behind `RLIMIT_FSIZE` with a default ceiling of 1 GiB, while the same options struct accepts a much larger `disk_size_gb` backed by a single growing qcow2 file on the host. When the qcow2 crosses the ceiling the monitor takes `SIGXFSZ`, and the shim installs no handler for it, so the process dies leaving no exit file and no crash record.

Evidence

• `disk.qcow2` stops at exactly `1073741824` bytes, and stops at the same byte with `disk_size_gb: 32`.
• Replacing the workload (an npm install) with plain `dd` reproduces it identically, so it is write volume, not the tool.
• 20 seconds after the death every further exec returns `Connection refused` — the VM is gone.
• The console log ends mid-normal-operation: no OOM kill, no panic, no kernel oops.
• Disk and memory were both healthy 5 s before death (`2371 used / 5644 free` MiB disk, `288 used / 1698 avail` MiB memory).

Why it matters

Any guest workload that writes more than a gibibyte hits this, which is ordinary for a package install or a build. Because the monitor leaves no record, it surfaces as an unrelated-looking transport error, so the time to diagnose is long.

Suggested fixes

1. Derive the file ceiling from the disk size the caller asked for, rather than a fixed default that can contradict it.
2. Handle `SIGXFSZ` in the shim and write a crash record, so the failure names itself instead of appearing as an h2 error.
3. Consider whether `exit_code: -1` should be distinguishable from a real exit status, since it currently means "no exit status was ever observed".

We worked around 1 by sizing the ceiling from the requested disk plus headroom on our side, but the default contradicting the accepted disk size looks worth fixing upstream.