chore(scaleway): the Terraform provider pin moves to 2.82.0, and the move measures what it exposed

[stephrobert]

The three exact pins in this repository move from 2.81.0 to 2.82.0: the
conformance fixture, examples/stacks/scaleway and examples/quickstart/scaleway.
The pin is exact so that moving it is an act, and this is that act.

What the release changes that could reach this emulator

Read from the release notes rather than met one failure at a time. 2.82.0 adds
an srn attribute to six products, three of which this emulator serves —
iam, vpcgw, ipam/flexibleip — plus fix(instance): infer project_id from server if not explicit, on a pack this repository does serve. Everything else is a new
product (autoscaling, datawarehouse) or a test chore.

What that turned out to mean, measured

srn is computed client-side, and the SDK that 2.82.0 bumps to asks the gateway
for the domain to build it with:

apiMetadata, err := s.client.GetAPIMetadata()   // GET /metadata
if err == nil {
    resp.setSRN(apiMetadata.Domain)
}

This emulator mounts no /metadata. One apply plus destroy of the
conformance fixture, recorded through feint proxy --record:

exchanges	539 over 100 distinct paths
GET /metadata	148 calls, every one answered 404
response codes	{200: 339, 404: 169, 204: 20, 201: 9, 202: 2}

148 of the run's 169 refusals are that one route. Nothing fails, because the
error is discarded — that is luck rather than a decision, and it is the shape of
#257 exactly.

So the answer is known rather than guessed. A read-only shot at a real
fr-par account, no resource created, 2026-09-14:

GET https://api.scaleway.com/metadata  ->  200
{"platform": "external", "partition": "scw", "domain": "scw.eu"}

It is not mounted in this PR, and the reason is the instrument rather than
the handler: Route.Operation must name an operation the drift scan finds, and
that scan walks api/<product>/<version> only, where GetAPIMetadata is not. A
route declaring it becomes an orphan, and all three baselines carry zero. That
decision is #776; docs/limits.md carries the measurement in the meantime.

The same shot recorded a second divergence, left to its own issue: the gateway
answers an unknown path with a Scaleway error document, where this emulator
answers net/http's plain-text 404 page not found.

Proven

• conformance:leg -- terraform: green, and the log names the client that
  answered — Installed scaleway/scaleway v2.82.0 (signed, key ID F5BF26CADF6F9614).
• conformance:leg -- scw-cli: green.
• conformance:quickstart: green — Scaleway brought up, second plan empty,
  taken down. That empty second plan is the assertion that catches a provider
  which gained an attribute the emulator does not serve.
• conformance:stacks: green — applied, second plan with no resource change,
  destroyed, on all three stacks.
• conformance:functional under incus-ovn, Scaleway half green: seven
  machines, each running and each really Running on the host, public paths,
  firewall rules, cross-VPC isolation, a restart through the API, stack taken
  down.
• mise run check and docs:check green; the generated client tables and the
  README safety banner regenerated (56 → 57 limits sections).

Two things about the registry, because one of them nearly fooled me

A false alarm: the registry answered 504 on 2.82.0's SHA256SUMS twice
running, which reads exactly like a broken release. Fifteen requests against
2.82.0 and fifteen against 2.81.0 as a witness came back 200/200 on both,
same byte size. One attempt at a remote service settles nothing.

A real outage, and not this release: the functional leg's Outscale half died
on Failed to install provider outscale/outscale v1.8.0[…] 504 Gateway Timeout returned from github.com. Fifteen requests for that asset came back 5 in 200
and 10 in 504, while both Scaleway assets were 15/15 the same minute. So the
Outscale half of that leg is not measured here, for a reason upstream of this
repository and unrelated to this pin — and CI may hit the same thing on any leg
that installs the Outscale provider.

Not in this change

mise run limits:check is red on 9 sections whose cited issues closed after
anyone last looked. It was red on main before this branch — verified by
stashing — and is a separate piece of work. The section added here carries its
own acknowledgement and is not among the nine.

🤖 Generated with Claude Code