fix(disks): eliminate DragonFly disklabel fan-out and md phantoms

[ncode]

Summary

The DragonFly disks/partitions probe had two independent problems, both confirmed on the nlab DragonFly guest while diagnosing the CI timeout:

1. Fan-out. dragonFlyDisklabelTargets blindly probed device + s1..s4 (plus the whole disk) for every device in kern.disks. The guest lists 132 devices (da0, vn0..3, md0..md126), so one discovery spawned ~660 disklabel processes, of which exactly one (disklabel da0s1) was productive.
2. Phantom disks. The 127 empty md memory disks leaked into the disks fact as phantom 9.77 MB disks (diskinfo /dev/md0 returns a positive size).

What changed (DragonFly only)

• dragonFlyPseudoDisk skips md/cd pseudo-devices (matched by driver class — name with trailing digits stripped) in both the disks list and the partition probe.
• dragonFlyDisklabelTargets now globs the slice nodes that actually exist (/dev/<dev>s<N>) instead of the fixed s1..s4, and drops the never-productive whole-disk target (the DragonFly label lives on the slice; disklabel da0 returns Operation not supported).
• currentDragonFlyPartitions takes s.glob. Unattached vn is handled by the existing size gate + empty glob; attached vn is kept.

Why DragonFly-only (ground truth from nlab)

OS	hw.disknames / kern.disks	Phantom leak?
DragonFly	md1..md126 da0 vn0..3 md0	Yes — 127 md phantoms, ~660 spawns
OpenBSD	sd0:<duid>,fd0:	No — rd0/cd0 not listed; fd0 errors out
NetBSD	sd0 dk0 dk1	No — md0/cd0 not listed; dk wedges already handled

OpenBSD/NetBSD hw.disknames is kernel-curated to attached instances, so pseudo-devices never enter the probe. Filtering rd there would be dead code — and harmful in a bsd.rd recovery boot where rd0 is the root disk. So this change does not touch their code paths; their facts are byte-identical by construction. Ruby Facter has no BSD disks resolver, so there is no parity constraint.

Validation (nlab DragonFly guest)

• Release gate: green (all facts present).
• disks = {da0} only — the 127 md phantoms are gone.
• partitions = /dev/da0s1a (/boot), s1b (swap), s1d (/) — unchanged.
• Discovery of disks partitions: 1m42s system time → 0.65s.
• disklabel spawns: ~660 → ~1.

Review

• OCR (--audience agent): 0 comments.
• Adversarial review: no correctness findings; one low test-gap (missing positive attached-vn test) was fixed.

Implements OpenSpec change optimize-bsd-disks-probe.