[smklein]

Fixes #273.

The recovery image removes six shared objects that libtopo and the topo
plugins loaded on Oxide hardware depend on. With omicron's in-progress
change to read disk locations from libtopo, installinator in the recovery
image fails its first snapshot ("empty topology") and is then killed by
ld.so.1 on a later one (librename.so.1: open failed, referenced by
libipcc_inv_persist). This stops elide-ing them from recovery-elide-v3.json.

Library	Size	Needed by	Without it, in the recovery image
librename.so.1	19.7 KiB	libipcc (inventory cache write, via oxhc); libdladm (libtopo → libuuid → libdlpi chain)	Process killed (librename_atomic_init); libuuid's dlpi_walk binding fails
libfmd_agent.so.1	24.5 KiB	pciebus.so, zen.so	Both fail to load; hc tree drops from 1226 to 409 nodes
libdiskstatus.so.1	99.1 KiB	disk.so	Fails to load; NVMe disk nodes missing
libjedec.so.1	309.0 KiB	dimm.so	Fails to load; DIMM nodes missing
libipmi.so.1	99.4 KiB	libtopo (ipmi_open)	The first plugin dlopen (pciebus) fails
libidspace.so.1	13.5 KiB	libvarpd (libtopo → libuuid → libdlpi chain)	The first plugin dlopen (pciebus) fails

Total is about 565 KiB, compared with a 149 MiB recovery zfs.img.

How the set was chosen (to be clear: by claude)

• Closure: ldd -Lr over libtopo and every 64-bit topo plugin on an
  installed Gimlet (helios-3.0.24348) gives 57 libraries. Nine of them are
  missing from the recovery image: the six above plus libsff, libscsi
  and libses.
• What loads on Oxide: the Gimlet, Cosmo and Metro maps use only
  oxhc, pciebus and fac_prov_libhp, and oxhc loads ufm, disk,
  dimm and zen. A 64-bit program taking repeated snapshots loaded
  exactly those seven plugins. libsff, libscsi and libses are needed
  only by nic.so and ses.so, which don't load on Oxide (only the
  generic pcibus module loads them), so they stay elided.
• Leave-one-out in recovery: on a Cosmo booted into the recovery image
  (same helios build), I added the six libraries and removed each one in
  turn, with the libipcc inventory cache absent, as during a mupdate. With
  all six: three full snapshots (1226 hc nodes), and the cache is written.
  Each omission fails as shown in the table.

libipmi and libidspace look optional at first: the hc tree still fills
in without them. But the first RTLD_NOW plugin dlopen fails, and later
ones succeed only because rtld treats the partially relocated libtopo as
already relocated. That leaves an unresolved lazy reference in place, which
is how librename turned into a fatal error rather than a failed dlopen.

[smklein]

I am in the process of testing this with oxidecomputer/omicron@58c912e

Basically:

• Recovery image construction prevents us from using libtopo instead of libdevinfo to monitor for disks in omicron #273 describes how I demo'd the issue atop Report each disk's chassis location in sled inventory omicron#11279 (installinator dies, mupdate fails, sad)
• oxidecomputer/omicron@58c912e is just "omicron 11279 + this change". The goal here: confirm that installinator now DOESN'T die when we include these shared libraries.

I'll report results when they're ready.

[smklein]

Mupdate succeeded with both omicron#11279 (using libtopo) + the omicron branch (oxidecomputer/omicron@58c912e) which adds all these shared libraries to the recovery image. This test was performed on london, mupdating sled 14 + 17 (failed without this change, succeeds with this change).

(If it matters: sled 14 is a gimlet, sled 17 is a cosmo. both seem update-able via installinator)

[citrus-it]

Otherwise LGTM, and I test built it and threw it on a bench gimlet out of interest.