egress-proxies-operator/docs/plans-executions/2026-09-18-1502-proxmox-provider.md

# Execution: proxmox, a Proxmox VE provider built from the provider contract

Pairs with [docs/plans/2026-09-18-1502-proxmox-provider.md](../plans/2026-09-18-1502-proxmox-provider.md).

## Status

- [x] Phase 0 — Plan file, branch, host recon, fixtures
- [x] Phase 1 — The spec (`docs/proxmox-provider-spec.md`)
- [x] Phase 2 — Config block, registration, package skeleton
- [x] Phase 3 — Extract `nocloud` and `keyedlock` from kvm2
- [x] Phase 4 — Seam, HTTP client, decorator, fake
- [x] Phase 5 — Get, ListByTag, task helpers
- [ ] Phase 6 — Create
- [ ] Phase 7 — Delete, Checker, CRD comment
- [ ] Phase 8 — Contract and observability sweep, integration test, testing.md
- [ ] Phase 9 — Deployment and documentation
- [ ] Phase 10 — Review, real-host validation, changelog, MR

## Phase 0 — Plan file, branch, host recon, fixtures

The plan was committed alone on `main` on 2026-09-18 (`708989c`, then
`2167819` retargeting it at PVE 9.2). `origin/main` had not moved in the
meantime, so the branch was cut straight from local `main`:

```bash
git fetch -q origin && git rev-list --left-right --count origin/main...main   # 0 2
git switch -c feat/proxmox-provider
```

The host is **pv1-1**, a single node (no cluster peers), `pve-manager/9.2.20`,
kernel `7.0.14-17-pve`, QEMU 11.0. The user prepared the token
(`~/.config/proxmox-token`, privsep off) and the CA (`~/.config/proxmox-ca.pem`)
before the session.

**TLS: 192.168.2.8 is not in the node certificate.** The first probe failed
with curl exit 60. The CA is right (`openssl s_client -CAfile` verifies the
chain), but the leaf's SANs are `127.0.0.1, ::1, localhost, 192.168.0.8,
pv1-1, pv1-1.home.lab`, and neither name resolves on the workstation.
192.168.0.8 and 192.168.2.8 present the same certificate fingerprint: the
node sits on `vmbr0` (192.168.0.8/24, untagged) with VLAN interfaces
`vmbr0.2` (192.168.2.8/24) and `vmbr0.4` (192.168.4.8/24). All recon kept
full verification by connecting to 192.168.2.8 while verifying the
certificate as `pv1-1.home.lab`:

```bash
curl -s --resolve pv1-1.home.lab:8006:192.168.2.8 --cacert ~/.config/proxmox-ca.pem \
  -H @<(printf 'Authorization: PVEAPIToken=%s\n' "$(tr -d '[:space:]' < ~/.config/proxmox-token)") \
  https://pv1-1.home.lab:8006/api2/json/version
# {"data":{"release":"9.2","version":"9.2.20","repoid":"49318c671b82f31e"}}
```

Every capture went through a scratch helper wrapping exactly that call. It
passes the token through a file descriptor so it never reaches argv or
output, and writes `{"status","reason","body"}` to
`internal/provider/proxmox/testdata/<name>.json`, refusing to save a body
that contains `PVEAPIToken`.

A real NoCloud seed was used instead of the plan's 1 KiB placeholder, so the
agent could be installed and F21 captured from a live guest. No ISO tool is
installed locally, so a throwaway module in the scratchpad reused the
repository's `github.com/kdomanski/iso9660 v0.4.0` from the module cache and
copied kvm2's `guestAgentScript` verbatim:

```bash
GOFLAGS=-mod=mod GOPROXY=off GOSUMDB=off go run . proxy-reconaaaaaaaaaaa proxy-reconaaaaaaaaaaa-seed.iso
# 45056 bytes, ISO 9660 'cidata'
```

Write experiments ran against VMID 199999 (`proxy-reconaaaaaaaaaaa`), a
clone of template 9001. Sequence: validation error; clone; resources, config
write and start while the clone lock was held; status and agent while
running bare; stop; config read; clone onto the taken VMID; tag tests;
upload twice plus a bad name; attach `ide2`, set cores and memory; three
resizes; start and agent poll; suspend and resume; a clone onto a storage
without image support; stop; destroy with `purge=1&destroy-unreferenced-disks=1`;
destroy again; ISO delete.

### Facts F1 to F23 (cited as Evidence by the spec)

- **F1.** Errors carry the text twice: in the HTTP reason phrase and as the
  body's `message`, which ends in `\n`. A 400 adds `errors`, a map from
  parameter to message. A 401 has the reason phrase only and an empty body.
  The client should use the trimmed body `message` and fall back to the
  reason phrase. Headers are ordinary CRLF.
- **F2.** Missing VM, identical for `GET config`, `GET status/current` and
  `DELETE`: `500 Configuration file 'nodes/pv1-1/qemu-server/999999.conf' does not exist`.
- **F3.** A start issued while the VM is clone-locked is **accepted** (200
  with a UPID). The task log reads `trying to acquire lock... OK` and the VM
  starts once the clone finishes. Setting a lock through the API was not
  tried.
- **F4.** Clone onto a taken VMID, synchronous:
  `500 unable to create VM 199999: config file already exists`.
- **F5.** Agent enabled but not running: `500 QEMU guest agent is not running`.
  Agent not enabled in the config: `500 No QEMU guest agent configured`.
  Stopped VM with `agent: 1`: `500 VM 199999 is not running`.
- **F6.** Not produced. `local-lvm` is thin and overprovisions, so
  `quota: absent`.
- **F7.** Config write while clone-locked: `500 VM is locked (clone)`.
- **F8.** `/cluster/resources?type=vm` entries carry `vmid, name, node,
  status, type, template, tags, maxcpu, maxmem, maxdisk, uptime` and more;
  `lock` appears only when set. Status values seen: `running`, `stopped`,
  and `unknown` during a clone. **Mid-clone the entry has no `name`**: only
  `type, vmid, id, node, lock: clone, status: unknown`.
- **F9.** Hyphenated tags are **accepted** and stored as given. Uppercase is
  folded (`Proxy_Operator` → `proxy_operator`), the separator is `;`, and
  characters such as a space or `!` give
  `400 … tags: invalid format - invalid characters in tag`. The datacenter
  has no `tag-style` or `user-tag-access` set.
- **F10.** `description` on clone is accepted and survives verbatim,
  newlines included. The 22-character `proxy-<base32>` name is accepted.
- **F11.** The clone's `meta` is copied from the template
  (`creation-qemu=11.0.3,ctime=1790145676`, identical to 9001), so it says
  nothing about when the clone was made. `written-at-create` is required.
- **F12.** Upload: multipart with field `content=iso` and the file part named
  `filename`, carrying the target file name. Returns a UPID for an
  `imgcopy` task. A name without `.iso` gives
  `400 … filename: wrong file extension`. **Re-uploading an existing name
  silently overwrites it** (the task log shows a plain `cp`).
  `Datastore.AllocateTemplate` suffices.
- **F13.** Deleting a volume needs `Datastore.Allocate` on
  `/storage/<id>`; without it the call gives
  `403 Permission check failed (/storage/local, Datastore.Allocate)`. With it
  (on `proxy-seeds`), the delete **runs as an `imgdel` task** and returns a
  UPID. Deleting a volume that no longer exists is also accepted and its
  task still ends `OK`; the log reads `disk image '…' does not exist`, then
  `Removed volume`. Seed cleanup is idempotent with no message matching.
- **F14.** Stopped: `status: stopped, qmpstatus: stopped`. Paused:
  `status: running, qmpstatus: paused` in `status/current`, while
  `/cluster/resources` shows plain `running`.
- **F15.** A clone onto a storage without image support is rejected up
  front (`400 … storage 'local' does not support vm images`) and leaves no
  VM. A clone failing mid-task was not observed.
- **F16.** Task status fields: `status` (`running`, `stopped`), `exitstatus`
  (`OK` or the error text), `type, id, node, user, tokenid, upid, starttime,
  pid, pstart`. A full clone of the 3 GiB template onto `local-lvm` finished
  within about two seconds.
- **F17.** The planned role, granted on `/`, covered clone, config, resize,
  start, stop, suspend, agent read, upload and destroy. **Missing:
  `Datastore.Allocate` for deleting the seed ISO.** `VM.Config.Cloudinit` is
  not needed for the `ide2` write. `SDN.Use` was held but never exercised,
  because no call rewrote `net0`.
- **F18.** HTTP/1.1 only: `curl --http2` negotiates `HTTP/1.1`.
- **F19.** Template 9001 (`debian-13-template`): `scsi0` on `local-lvm`,
  `boot: order=scsi0`, `ide2` free, `agent: 1`, OVMF, q35, `net0` untagged on
  `vmbr0`. **Attaching `ide2` makes Proxmox append it to the boot order**
  (`order=scsi0;ide2`). Template 9000 (Ubuntu) is unsuitable: it has a
  Proxmox cloud-init drive on `ide2` and VLAN tag 4 on `net0`.
- **F20.** Resize **runs as a task** (returns a UPID). The same size is an
  `OK` no-op; shrinking fails inside the task with `exitstatus`
  `shrinking disks is not supported`.
- **F21.** Agent response: `result[]` of `{name, ip-addresses[{ip-address,
  ip-address-type}]}`. The live guest reported `lo` and `ens18` with
  `192.168.0.150`, 50 seconds after start, via DHCP on untagged `vmbr0`.
  This proves the NoCloud-seed-on-`ide2` route end to end on this host.
- **F22.** A wrong secret and an unknown token id both give
  `401 Authentication failed!`; no credentials gives `401 No ticket`.
- **F23.** Uploads are staged in `/var/tmp/pveupload-*` and then copied to
  `/var/lib/vz/template/iso/`, so the tmpfs `/tmp` is not involved.

### Corrections to the plan that Phase 1 must fold into the spec

1. **Check `lock` before the name.** A mid-clone VM has no name in
   `/cluster/resources` (F8), so name-first logic turns a clone in progress
   into NotFound. Get, Delete and Create's lookup must treat a locked or
   nameless entry at our VMID as in-progress (Transient or wait), never
   absent.
2. **Inspect a taken VMID before probing forward.** After a crash mid-clone,
   the retry's name lookup misses the nameless VM, the clone onto the
   deterministic VMID fails with F4, and forward probing would clone a
   duplicate. On F4, read that VMID: `lock: clone` means wait for it, and our
   UID in the description means adopt it. Probe forward only for a VM that
   is demonstrably someone else's.
3. **A start queues behind a lock** instead of failing (F3).
4. **Resize is asynchronous** (F20); wait on its task.
5. **Drop delete-before-upload** for the seed: upload overwrites (F12).
   **Seed deletion is a task** (F13): wait on it. A repeat delete succeeds,
   so Delete needs no absent-volume matcher.
6. **Get reads `status/current`**, not only `/cluster/resources`, to tell
   paused from running (F14).
7. **Tag rationale**: hyphens are accepted on 9.2.20 (F9). `proxy_operator`
   stays, since it is valid everywhere, but the claim that hyphens are
   rejected is removed.
8. **The boot-disk logic must skip `ide2`**, which Proxmox adds to the boot
   order (F19). The plan's rule already excludes it; add a test with
   `order=scsi0;ide2`.
9. **Three agent messages, not one** (F5): `QEMU guest agent is not running`
   (booting, or the install failed), `No QEMU guest agent configured` (our
   `agent=1` was removed), and `VM <vmid> is not running` (stopped between
   the status read and the agent call). The spec decides each class; none
   may become NotFound.
10. **Internet on the proxy network is a hard prerequisite**, twice over: the
    vendor-data agent install needs a package mirror, and the proxy itself
    needs egress. The guide says so, and recommends baking
    `qemu-guest-agent` into the template with `installGuestAgent: false` to
    drop the mirror dependency from IP discovery.

### Decisions made with the user in Phase 0

- **Seeds go to a dedicated directory storage**, `proxy-seeds`, with ISO
  content only. The token gets `Datastore.Allocate` there and nowhere else,
  so it can delete seeds but never backups or other ISOs on `local`. The
  sample config's `isoStorage` becomes `proxy-seeds`. The guide must say that
  an ACL on a more specific path replaces the inherited one for that user,
  so the role granted on `/storage/proxy-seeds` must repeat
  `Datastore.AllocateTemplate`, `AllocateSpace` and `Audit` alongside
  `Allocate`.
- **Proxies land on a VLAN** set in config. `ProxmoxConfig` gains an
  optional `vlanTag` (1 to 4094), applied when Create rewrites `net0`
  together with the optional bridge override, keeping the MAC. Unset leaves
  `net0` as cloned. This is the first call that exercises `SDN.Use` (F17).
- **The API URL is `https://192.168.0.8:8006`**, the address in the node
  certificate, so `caFile` verifies with no provider change. No
  `tlsServerName` option. The operator's cluster must be able to reach
  192.168.0.x.

### Seed storage and VLAN 4

The user created `proxy-seeds` (`dir`, ISO content only), granted the token
`Datastore.Allocate`, `AllocateSpace`, `AllocateTemplate` and `Audit` on
`/storage/proxy-seeds` through a role `ProxyOperatorSeeds`, and freed the
leftover recon ISO on `local`. An upload to `proxy-seeds` and two deletes of
a throwaway ISO there produced `content_delete_response.json`,
`content_delete_again_response.json` and `task_status_volume_missing.json`.

The user gave the VLAN as "vlan 4, bridge: vmbr0.4@vmbr0". That is the
host's own VLAN interface name as `ip link` prints it; a VM NIC cannot
attach to it because it is not a bridge. On the VLAN-aware `vmbr0` the
Proxmox form is `bridge=vmbr0,tag=4`, which is what template 9000 uses, so
the provider's `vlanTag: 4` appends `tag=4` to the cloned `net0`.

A second clone of 9001 at VMID 199999 got its seed from `proxy-seeds` and
`net0` rewritten from `virtio=BC:24:11:EE:48:83,bridge=vmbr0` to
`virtio=BC:24:11:EE:48:83,bridge=vmbr0,tag=4`. **The write succeeded, so the
token's `SDN.Use` on `/` covers a VLAN tag on `vmbr0`**
(`set_config_vlan_response.json`).

**The first VLAN 4 attempt never produced an agent** in seven minutes of
polling (every 13 seconds, 33 polls). The only signal was the agent call, so
the cause could not be seen from the API. A debug seed added a `recon` user
(user-data only, with a new instance id so cloud-init re-ran), and the user
logged in on the console: DHCP worked, but VLAN 4 had no route to the
internet because of a routing misconfiguration on the Proxmox host, so the
vendor-data `apt-get install qemu-guest-agent` could not run. The user fixed
the routing.

**Rerun from scratch after the fix**, same sequence the provider will use
(destroy and seed delete first, then a fresh clone and seed, `instance-id`
equal to the name):

| Step | Elapsed |
|---|---|
| clone task done | 4 s |
| seed upload task done | 4 s |
| config (`ide2`, `net0` with `tag=4`, tag) | 4 s |
| start task done | 5 s |
| agent reports `192.168.4.147` | 47 s after start |

The guest was reachable from the workstation (ping, TCP 22), as the contract
requires of a proxy address. Stopping it gave the F5 stopped-VM fixture;
destroy and seed delete then left the host with only its original VMs 100,
9000 and 9001, no ISOs on `local` or `proxy-seeds`, and no `local-lvm`
volumes of 199999. The capture is `agent_network_get_interfaces_vlan.json`.

### Worth noting

- The token file is mode 664; 600 would be tighter.
- Fixtures: 48 files, all valid JSON, none containing `PVEAPIToken`.
  Several names differ from the plan's list because what they captured
  differed from what was expected: `start_while_clone_locked_response.json`
  (the start succeeded), `task_status_stopped_error.json` (from the shrink,
  not a start), `upload_overwrite_response.json` (success, not an error),
  `resize_same_size_response.json` and `resize_shrink_response.json`
  (UPIDs), `err_forbidden_volume_delete.json` (the 403 from F13).
  `err_no_space.json` was not produced (F6).
- The host is back to its original state; see the table above.
- The scratch seed builder (`seedgen` in the session scratchpad) took the
  user-data file and instance id as arguments by the end, for the debug
  login. It is not part of the repository; Phase 3's `nocloud` package
  replaces it.
- Time to IP was about 50 s on both networks, with the agent installed at
  first boot. Budgets in the plan (running wait 60 s for the VM state, the
  agent discovered by `Get` polling) are consistent with this.

## Phase 1 — The spec

Wrote [docs/proxmox-provider-spec.md](../proxmox-provider-spec.md), all 24
entries of the template, from the plan, the contract and the Phase 0 facts.
The ten Phase 0 corrections are folded in, and the plan now opens with a
note that the spec wins where the two disagree.

Judgment calls made here, each recorded in its entry:

- **Create's allocation loop** (entry 2) runs an ownership check on any
  taken VMID: wait out a lock or a nameless entry, then adopt when the name
  and UID are ours, skip when the name differs, refuse as Permanent when the
  name is ours but the mark is not. This is the fix for the duplicate-VM bug.
- **Get's check order** (entry 9): lock, then name and mark from the config,
  then `status/current`, then the agent. `lock: suspended` alone maps to
  Stopped, since a hibernated VM never resumes by itself.
- **Agent messages** (entry 7): `QEMU guest agent is not running` is
  Provisioning, `VM <vmid> is not running` is Transient, and
  `No QEMU guest agent configured` is **Stopped**, so the reconciler replaces
  a VM that can never report an address instead of latching the Proxy.
- **Templates are checked before cloning** (entry 4): not a template, or a
  Proxmox cloud-init drive in any slot, is Permanent. This makes the "no
  cloud-init drive" prerequisite enforceable instead of a silent failure.
- **`isoStorage` is required**, with no default (entry 15), because the token
  must hold `Datastore.Allocate` on it.
- **TLS verification failures are Permanent**, with a Checker reason
  `CertificateInvalid` (entries 13, 16), since the fix is the URL or CA file.
- **The CA is loaded lazily and cached**; the token is read per request
  (entry 14).

Checks, as the plan's Verify clause asks:

```bash
grep -c '^## [0-9]*\. ' docs/proxmox-provider-spec.md     # 24
grep -n '^\*\*Answer:\*\*$' docs/proxmox-provider-spec.md # nothing
grep -c '—' docs/proxmox-provider-spec.md                 # 0
```

plus a script comparing every `` `<axis>`: `<profile>` `` pair against the
contract's §4.20 table (no mismatches, every axis answered) and a list of
the F numbers cited (all 23).

Worth noting: the first pass stated two things as fact that recon could not
show and was corrected before commit. That an unreachable node's guests stay
listed as `unknown` is Proxmox's documented design, unverifiable on a single
node; entry 13 now says so. And the matcher for `VM <vmid> is not running`
is a whole-message regexp, because ` is not running` alone also matches
`QEMU guest agent is not running`.

## Phase 2 — Config block, registration, package skeleton

Added the `proxmox:` block (`internal/provider/config_proxmox.go`), wired it
into `config.go` (the field, `setBlocks`, the allow-list and the dispatch:
three `"proxmox"` sites) and registered `"proxmox": proxmox.New` in
`cmd/main.go`, so the two type lists changed in one commit. The package
`internal/provider/proxmox` now holds the seam interface and its data types
(`api.go`), `id.go`, `mark.go`, `state.go`, `errors.go` and the constructor
(`proxmox.go`). The four contract methods are stubs that answer
`ErrTransient`, so a Proxy on a configured `proxmox` entry retries instead of
latching until Phases 5 to 7 land. `New` passes a nil `api` until Phase 4
builds the HTTP client.

Every test that depends on Proxmox's wording or JSON shape reads the Phase 0
captures through `fixtures_test.go` (`readCapture`, `loadAPIError`,
`decodeData`), so no message string is repeated in a test. The one exception
is `TestClassify_table`'s synthetic rows (a 404, 502, 595, 596, a storage
message), for statuses recon never produced.

Deviations from the plan's Phase 2 list, both making it smaller:

- **No temporary `locks.go`.** Only Create and Delete take the per-name lock,
  and they arrive after Phase 3 extracts `internal/provider/keyedlock`, so
  copying kvm2's lock now would only be deleted next phase.
- **No `provider.Checker` assertion yet.** `cmd/main.go` type-asserts the
  Checker and would call a stub; the assertion comes with the real `Check`
  in Phase 7.

Added beyond the plan: `api_test.go` decodes the captured responses into the
seam's types, so a misspelled JSON tag fails now rather than in Phase 5, and
`TestParseDescription_fromClone` reads the mark back from the description of
the VM cloned in Phase 0.

A mutation check confirmed the suite can fail. Matching `VM <vmid> is not
running` as a plain suffix (the first spec draft's mistake) failed
`TestAPIError_matchers` on the agent-not-running capture; treating a paused
VM as running failed `TestMapState_table` and `TestMapState_fromCaptures`.
Both mutations were reverted and the files compared identical to backups.

```bash
gofmt -w cmd/main.go            # realigned the constructor map for the longer key
grep -c '"proxmox"' internal/provider/config.go                     # 3
go build ./... && go vet ./...
go test -race -count=1 ./internal/provider/... ./cmd/...            # all ok
go test -short -count=1 ./...                                       # all ok
go test -count=1 -v ./internal/provider/proxmox/ | grep -c '^=== RUN' # 152
```

Worth noting: the `validProxmoxYAML` constant is the shape Phase 9's commented
sample entry should take, with the user's values (`192.168.0.8`, `pv1-1`,
`proxy-seeds`, `vlanTag: 4`, template 9001).

## Phase 3 — Extract `nocloud` and `keyedlock` from kvm2

kvm2's `seed.go` and `seed_test.go` moved to `internal/provider/nocloud`
with `git mv`, so history follows them. The package exports `BuildSeedISO`,
`BuildVendorData`, `GuestAgentScript`, `SeedLabel` and a new `ErrEmptyName`
sentinel in place of kvm2's `permanentf`. kvm2's `keyedLocks` became
`internal/provider/keyedlock` (`Locks`, `New`, `Acquire`), and its three
`TestKeyedLocks_*` tests moved with it under their old names. kvm2 now
imports both packages, and its `create.go` still wraps every seed-build error
with `permanentf`, so kvm2's behaviour is unchanged. proxmox gained the
`nameLocks` field and `lockName`.

The test counts match the plan's Verify clause: kvm2 dropped from 465 to 447
`=== RUN` entries, which is the 12 moved tests plus the 6 subtests of
`TestSeedISO_userDataByteForByte`. `nocloud` (15) and `keyedlock` (3) carry
exactly those 18.

```bash
git mv internal/provider/kvm2/seed.go internal/provider/nocloud/seed.go
git mv internal/provider/kvm2/seed_test.go internal/provider/nocloud/seed_test.go
go test -count=1 -v ./internal/provider/kvm2/ | grep -c '^=== RUN'   # 465 before, 447 after
grep -rn 'iso9660' internal/provider/kvm2/                          # nothing
go build ./... && go vet ./... && go vet -tags=integration ./...
go test -race -count=1 ./internal/provider/...                      # all ok, run three times
go test -short -count=1 ./...                                       # all ok
```

**Pulled forward from Phase 6 so proxmox imports `nocloud` now**, as the
"Done when" line requires: `seed.go` with `seedISOName`, `seedVolID` and
`buildSeed`, the pure part of the seed step. `uploadSeed` still belongs to
Phase 6, which calls `buildSeed`. The new tests are `TestSeedNames_table`,
`TestBuildSeed_table` (4 rows),
`TestBuildSeed_emptyNameIsPermanentWithoutPayload`,
`TestBuildSeed_stagingFailureIsTransient` and
`TestLockName_givesUpWithContextAsTransient`. proxmox went from 152 to 161
test entries.

**proxmox does not copy kvm2's open review finding S5.1**
(`docs/reviews/2026-09-17-kvm2-provider-review/sweep.md`). In kvm2, a
staging failure such as a missing `/tmp` or a full emptyDir becomes
Permanent, and that latches the Proxy. `buildSeed` makes only
`ErrEmptyName` Permanent and wraps any other error with `%w`, so
`classify` reads it as Transient. The sentinel makes the same three-line
fix possible in kvm2, but this phase leaves kvm2's behaviour unchanged.

Worth noting:

- The older `kvm` provider has its own `seed.go` with a different
  signature (`buildSeedISO` returns `[]byte`). The plan scopes the
  extraction to kvm2, so `kvm` is untouched.
- `GuestAgentScript` now says the operator reads the address "from the
  hypervisor" instead of "through libvirt", in both the Go comment and the
  script's comment line inside the guest. The script's commands are
  unchanged.
- The docs now name the new packages: `kvm2-provider.md`,
  `kvm2-provider-spec.md` entry 19, the kvm2 columns and section of
  `providers-existing.md`, plus a "Shared provider helpers" row in the
  `architecture.md` component table. `docs/reviews/` keeps the old names,
  because it records the code as it was reviewed.
- `TestCheck_hostsCheckedConcurrently` in kvm2 failed once under the full
  `-race` suite (latency 199.96 ms against a 200 ms floor). It passed 20
  runs out of 20 in isolation, with and without this phase's changes. The
  test starts its 200 ms release timers before `Check` starts its clock, so
  under load the measured latency can come in just under 200 ms. The flake
  predates this phase, and the test is not touched here.

## Phase 4 — Seam, HTTP client, decorator, fake

`client.go` holds `httpAPI`, the thin `net/http` adapter for all sixteen
seam methods. It reads the token file on every request, turns any non-2xx
answer into `*apiError` from the body's `message` or the reason phrase
(F1), and wraps its transport in `tracing.RestConfigWrapper`. `observe.go`
is the logging decorator: one V(1) `proxmox api call` line per call,
detail at V(2), and no span, because the transport already traces each
request. `New` now builds `observe(newHTTPAPI(...))` instead of passing a
nil `api`, still without touching the network or the token file.
`fake_test.go` holds `fakeAPI`, which Phases 5 to 7 build on.

Where this phase departs from the plan's Phase 4 text:

- **The `http.Client` is built on first use, not in `newHTTPAPI`.** Spec
  entry 14 wins over the plan: the CA is loaded lazily, a missing CA file
  is Transient and retried until the file lands, one without a certificate
  is Permanent, and the pool is cached only once it loads. A transport's
  `RootCAs` cannot change after first use, so the whole client waits for
  the CA (`TestClient_caFileLoadedLazily`).
- **The upload is not an `io.Pipe` with a goroutine.** The multipart
  framing (the `content=iso` field, the `filename` part header and the
  closing boundary) is rendered into two small buffers, and the body is
  `io.MultiReader(head, io.LimitReader(r, size), tail)` with an exact
  `Content-Length`. The seed is still streamed and never held by the
  client. There is also no goroutine to leak, and the request carries a
  length, as Phase 0's `curl -F` upload did. A chunked upload to pveproxy
  was never tried. `TestClient_uploadStreamsMultipart` proves the
  streaming with a reader that stalls after its first 256 KiB until the
  server has seen file bytes.
- **The fake's `DeleteVolume` tolerates a missing volume**, as spec
  correction 5 and F13 say, instead of the plan's error.
- **Added:** secret redaction in `apiError` (contract §7), because
  otherwise a server that echoes the header would put the secret into
  `Proxy` status; an `errCredentials` sentinel on token-file errors, so
  Phase 7's Checker can report `CredentialsUnavailable`; and the V(1)
  identifiers `newid` and `name` on a clone, `filename` on an upload and
  `volid` on a volume delete. Spec entry 20 now lists them.

**The volid's `/` travels as `%2F`.** Neither the plan nor the spec says
how `{volid}` is encoded in the content-delete path. The Phase 0 session's
transcript shows every delete went through a helper encoding the whole
volid:

```bash
enc() { python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1], safe=''))" "$1"; }
curl ... -X DELETE ".../storage/proxy-seeds/content/$(enc 'proxy-seeds:iso/proxy-reconaaaaaaaaaaa-seed.iso')"
```

So `apiPath` runs `url.PathEscape` over each segment
(`TestClient_volIDSlashIsEscaped`).

`net/http` cannot send a custom reason phrase, so the client tests replay
every capture byte for byte through a hijacked connection (`serveCapture`).
`TestClient_apiErrorFromStatusReasonAndBody` globs `testdata/err_*.json`
and skips `err_tag_hyphen.json`, which is a 200 despite its name: hyphenated
tags were accepted (F9).

Tests beyond the plan's list: `TestClient_methodPathAndParams` (the method,
path, query and form of every seam method, with a reflection check that no
method lacks a row), `TestClient_cloneOmitsEmptyStorageAndTarget`,
`TestClient_volIDSlashIsEscaped`, `TestClient_caFileLoadedLazily`,
`TestClient_x509FailureIsPermanent` (from spec entry 21),
`TestDecodeConfig_scalarsAsStrings`, `TestReadToken_trimsTrailingNewline`,
`TestObserve_startsNoSpan`, `TestObserve_identifiersAtV1`,
`TestNew_wiresObservedHTTPClient`, and six `TestFakeAPI_*` tests.
`TestFakeAPI_errorsMatchCaptures` compares the fake's error texts with the
captures, so every test that goes through the fake exercises the real
wording. Texts that recon never produced start with `fake:`.

A mutation check confirmed the new tests can fail: dropping `redact`,
buffering the upload with `io.ReadAll`, and logging V(2) detail at V(1)
each failed their test. Removing only the empty `TLSNextProto` map did
**not** fail `TestClient_http1Only`. A transport with a custom
`TLSClientConfig` and `DialContext` never offers h2 unless
`ForceAttemptHTTP2` is set, so the map is a second guard. With both
guards removed the test failed with `client spoke HTTP/2`. Every mutation
was reverted, and the files were compared identical to backups.

```bash
go test -race -count=2 ./internal/provider/proxmox/          # ok, run three times
grep -n 'ReadAll' internal/provider/proxmox/client.go        # only the 64 KiB error-body read
grep -n 'context.Background()' internal/provider/proxmox/*.go | grep -v _test.go   # nothing
go build ./... && go vet ./... && go vet -tags=integration ./...
go test -short -count=1 ./...                                # all ok
go test -count=1 -v ./internal/provider/proxmox/ | grep -c '^=== RUN'   # 161 before, 289 after
```

Worth noting:

- `bin/golangci-lint run ./internal/provider/proxmox/...` panics inside
  staticcheck's IR builder (`unexpected expr: *ast.KeyValueExpr` in the
  Go 1.27 standard library's `poll` package). This is a toolchain problem,
  not this code. With `--disable staticcheck` the new files' findings are
  goconst on form keys (kvm2 carries 50 of the same) and two unparam hits
  on the fake's `addTemplate` and `resourceByID`, which Phases 5 and 6 will
  call with other VMIDs. The mapsloop and gocyclo findings in `proxmox.go`
  and `api_test.go` date from Phase 2.
- Nothing in this phase talked to the host. The client has only met
  `httptest` servers replaying the captures. The first real request is
  Phase 10's integration run, which is where the `Content-Length` upload
  and the `%2F` volid get proven against pveproxy.
- The fake refuses to destroy a running VM with an uncaptured
  `fake: VM <vmid> is running - destroy failed`, so Phase 7's
  stop-before-destroy order is enforced by the fake even though the real
  text is unknown.

## Phase 5 — Get, ListByTag, task helpers

`task.go` holds `parseUPID`, a shared `pollUntil` loop, `waitTask` and
`waitUnlocked`. `get.go` holds `Get`, the guest-agent read and `resolve`
(one resources call, QEMU entries only, which Delete will reuse). `list.go`
holds `ListByTag` and its candidate rule. The Get and ListByTag stubs are
gone, and `TestStubs_answerTransient` now covers only Create and Delete.

The spec overrides the plan in the places the Phase 4 notes listed. Get
checks in entry 9's order, so a lock or `status: unknown` in the resources
entry is Transient before the name is ever read, and a running Get makes
four calls (`TestGet_runningMakesFourCalls`). The three agent messages
follow entry 7. The poll interval is `budgets.poll`, and a failed task logs
a V(1) `proxmox task failed` line with `upid` and `exitstatus`.

Judgment calls, most of them spec gaps:

- **`WARNINGS: <n>` counts as a successful task.** Proxmox reports a task
  that finished but logged warnings with that exitstatus and treats it as
  finished, not failed. Under the spec's old "anything but `OK` is
  Transient" rule, a start that logged a warning would make Create retry
  forever. Recon never produced one, so this row relies on how Proxmox is
  known to behave, not on a capture. The spec's entry 10 text and entry 13
  table now say so.
- **A `suspended` lock still goes through the name and mark checks.** The
  entry 9 table puts "lock suspended is Stopped" before the config rows.
  Taken literally, a hibernated VM that reused our VMID would be reported
  as ours and Stopped, and Delete would then skip it as foreign in a loop.
  Get therefore reads the config, which it needs anyway for UID and
  CreatedAt, applies the name and mark checks, and only then answers
  Stopped. It skips `status/current` and the agent. Entry 9 now says this.
- **`TestGet_twoVMsWithOurNameIsPermanent` became
  `TestGet_otherVMWithOurNameIsIgnored`.** Spec entries 2 and 9 address Get
  by VMID, and only Create refuses a duplicate name. Get answering
  Permanent would latch a working Proxy as Failed because of a VM someone
  else created.
- **`TestGet_unmanagedDescriptionHasEmptyUID` tests a mark with no `uid`
  line.** The plan's name read literally ("unmanaged") would contradict
  "no mark is NotFound".
- **Agent errors outside the entry 7 table are forced to Transient**, even
  a 403 or a `does not exist`. Otherwise `classify` would make them
  Permanent or NotFound, and the spec says "anything else: Transient".
- **`waitUnlocked` returns the unlocked config**, which saves Create's
  ownership check one read. When the VM disappears while it waits (a
  failed clone), it returns the does-not-exist error and lets Create
  decide.

Pulled forward and left behind:

- **`provider_test.go` now exists** with `newFixture`, `testBudgets`,
  `addOurVM` and `assertProviderError`, because every Get and ListByTag
  test needs them. The plan created this file in Phase 8. Phase 8 adds the
  contract sweep to it.
- **`waitRunning` is left to Phase 6.** The plan's package layout puts it
  in `task.go`, but Phase 5 lists no test for it and only Create uses it.
  It should be built on `pollUntil`.
- **The fake gained two things.** `fakeVM.kind` makes an entry listed as
  `lxc` invisible to the qemu endpoints. `hookOn` is a general per-method
  hook, and `failWith` is now built on it. Tests use `hookOn` (through
  `mutateOn` in `get_test.go`) to change a VM between two of Get's calls.

A mutation check confirmed the new tests can fail. Each of these failed a
test: classifying agent errors instead of forcing Transient; dropping the
config lock check; dropping the `suspended` special case; ignoring an ended
ctx in ListByTag; treating `WARNINGS` as failure; and not mapping
`No QEMU guest agent configured` to Stopped. Removing the resources-entry
lock and `unknown` checks together failed only
`TestGet_unknownStatusIsTransient`. The fake reports a lock in the config
as well, so the config check catches every lock row, including a mid-clone
VM. The resources check stays because the spec orders it first and
because it spares a config read. Every mutation was reverted, and the
files were compared identical to backups.

```bash
go test -race -count=2 ./internal/provider/proxmox/                     # ok
grep -n 'not implemented' internal/provider/proxmox/*.go | grep -v _test # Create and Delete only
go build ./... && go vet ./... && go vet -tags=integration ./...
go test -race -count=1 ./internal/provider/... ./cmd/...                 # all ok
go test -short -count=1 ./...                                            # all ok
go test -count=1 -v ./internal/provider/proxmox/ | grep -c '^=== RUN'   # 289 before, 393 after
```

Lint ran with staticcheck disabled, because staticcheck panics on
Go 1.27:

```bash
bin/golangci-lint run --disable staticcheck ./internal/provider/proxmox/...
```

Worth noting:

- The lint findings in the new files are goconst on test strings and
  logcheck on `get`, `agentAddress` and `waitTask` taking both a ctx and a
  logger. That shape copies kvm2, which has five of the same findings. The
  errcheck hits on `assertGetFails` were fixed by returning the error text.
  The two unparam hits and the mapsloop hits predate this phase.
- `waitUnlocked` waits on any lock, `suspended` included. Phase 6's
  ownership check should decide whether a hibernated VM at our VMID is
  worth waiting out the clone budget for, or should be refused at once.
- Nothing in this phase talked to the host. Every Proxmox message a test
  relies on comes from `testdata/` or from fake errors that
  `TestFakeAPI_errorsMatchCaptures` ties to the captures.