feat: VM provisioning, and reboot_host on ESXi

create_vm_from_cloud_init takes the same qcow2/raw cloud images netOrk
offers for Proxmox. ESXi can neither boot nor download them, so the
driver does both: download with checksum check and one retry, convert
with qemu-img to a streamOptimized VMDK (cached by URL), import through
a minimal OVF descriptor over NFC, pin requested MACs, grow the disk,
and attach a NoCloud seed ISO placed next to the VM's files. NoCloud
rather than guestinfo because it needs nothing in the guest; the
user-data installs open-vm-tools, which the driver declares as its
guest agent. A failure after the import removes the VM again.

Placement is a pure decision over inventory rows: a connected host
outside maintenance mode that sees the datastore and every port group,
with the resource pool and VM folder found by walking up to the
datacenter -- one path for a standalone host and for a vCenter.

Two faults vcsim surfaced and the tests now pin: a chunked upload body
next to a Content-Length is refused with 500, so the disk goes up as a
sized file object that also reports lease progress; and a device edit
replaces the device as sent, so disk and NIC edits start from the live
objects, backing included (vcsim panicked on a disk without one).

destroy_vm, get_vm_status, get_network_targets (port groups with their
fixed VLAN) and get_image_storages complete the contract.

reboot_host on ESXi uses RebootHost_Task and refuses outside
maintenance mode: force=True would cut power to running VMs.

Tested against vcsim in ESXi and vCenter mode, end to end.

README.md

@@ -30,7 +30,10 @@ hardware**: see [Harvesting fixtures](#harvesting-fixtures).
 | `get_vm_storage_pools` | ✅ | ✅ | datastores |
 | `get_virtual_networks` | ✅ | ✅ (+ dvPortgroups) | port groups |
 | `get_device_warnings` | ✅ | ✅ | raw `{code, meta}` |
-| VM provisioning, VIBs, updates, host reboot | — | — | out of scope for v1 |
+| `reboot_host` | ✅ (maintenance mode only) | — | `RebootHost_Task` |
+| `create_vm_from_cloud_init`, `destroy_vm`, `get_vm_status` | ✅ | ✅ | see [Provisioning](#provisioning) |
+| `get_network_targets`, `get_image_storages` | ✅ | ✅ | port groups, datastores |
+| VIBs, updates | — | — | not yet |
 
 Unverified assumptions, to be checked against real hardware:
 
@@ -44,6 +47,9 @@ Unverified assumptions, to be checked against real hardware:
 - An account that may read the inventory. For power and snapshot actions it
   also needs *Virtual machine → Interaction → Power on/off/Reset/Suspend* and
   *Virtual machine → Snapshot management*.
+- For provisioning: `qemu-img` (package `qemu-utils`) where the driver runs,
+  and HTTPS from there to every ESXi host that may receive a VM -- through a
+  vCenter the disk upload goes straight to the host, not via the vCenter.
 - **A paid license for any write.** On the free vSphere Hypervisor license
   the API is read-only; the driver reports `vmware_api_read_only` and turns
   the refusal into a readable error.
@@ -69,7 +75,9 @@ driver.close()
 ```
 
 `optional_args`: `port` (default 443), `verify_ssl` / `ssl_verify` (default
-`True`; ESXi ships a self-signed certificate). Other keys are ignored.
+`True`; ESXi ships a self-signed certificate), `image_cache_dir` (where
+converted cloud images are kept; default a directory under the system temp
+dir). Other keys are ignored.
 
 VMs are addressed by name, by `vmid`, or by MoRef (`vm-42`). A name shared
 by two VMs is refused rather than guessed.
@@ -94,6 +102,32 @@ tools/sanitize.py tools/harvest-out/esxi8-dell.json > tests/fixtures/esxi8-dell.
 the ones the drivers read. `tools/harvest-out/` is gitignored. Read the
 sanitised file before committing it.
 
+## Provisioning
+
+`create_vm_from_cloud_init` takes the same cloud images netOrk's catalog
+offers for Proxmox (qcow2/raw):
+
+1. The image is downloaded where the driver runs, its checksum verified (one
+   retry), and converted with `qemu-img` to a streamOptimized VMDK. The VMDK
+   is cached by URL in `image_cache_dir`; the download is not kept.
+2. A host is chosen that is connected, not in maintenance mode, and sees the
+   datastore and every requested port group (the named datastore, or the one
+   with the most free space).
+3. The VM is created from a minimal OVF descriptor (PVSCSI disk, VMXNET3
+   NICs) and the disk streamed in over NFC.
+4. Requested MACs are pinned, the disk grown to `disk_resize_gb`, and a
+   NoCloud seed ISO (user-data, meta-data, network-config) uploaded next to
+   the VM's files and attached as a CD-ROM on a new SATA controller.
+5. The VM is powered on. Any failure after step 3 removes the VM again.
+
+A port group fixes its VLAN, so `get_network_targets` reports each one with
+`kind="portgroup"`, `vlan_aware=False` and its `fixed_vlan_tag`; a NIC asking
+for a different `vlan_tag` is refused. The guest agent netOrk installs is
+`open-vm-tools` (`GUEST_AGENT_PACKAGES`), which reports the IP address back.
+
+Unverified until #305: that each distribution's cloud kernel carries the
+PVSCSI and VMXNET3 drivers.
+
 ## Design notes
 
 **One seam.** Every read goes through `Inventory.collect(type, paths)`, a