v0.4.0

Added

• Terraform drives Outscale. init, validate, plan, apply, a second
  empty plan and destroy, against the published provider outscale/outscale
  v1.7.0. Until now everything this pack claimed was proven by oapi-cli alone,
  and the provider walks paths no CLI does. Writing the fixture was enough to
  find six defects, each listed below.
• Volumes. Create, read, update, delete, link and unlink. A volume grows and
  refuses to shrink — a filesystem does not survive its disk getting smaller —
  and a linked volume refuses to go, which is what a client destroying in the
  wrong order needs in order to retry. Snapshots stay declined: there are no
  bytes behind an emulated volume, so restoring one would produce a disk holding
  nothing.
• Tags on every resource that carries them, sorted, because their order is a
  permanent Terraform diff waiting to happen. And ReadVmsState, the
  lightweight view a client polls.
• 74 of 104 routes are proven by a real client, up from 69 of 93. 23 more
  are probed only: the protocol holds and the behaviour is unproven.

Fixed

• terraform apply died on the first machine. The catalogue published no
  ProductCodes, so the provider called ReadAdminPassword on every Vm it read
  back, Linux included — it is a Windows call, and an absent list reads as
  "unknown". That route did not exist. Images and Vms publish the Linux code
  now, and the call answers an empty password, never a generated one: a made-up
  credential is one a client could try to use.
• Every terraform destroy crashed the provider outright — "Plugin did not
  respond", reproduced with the published, signed provider v1.7.0. DeleteVms
  removed the record, and the provider answers a delete by polling ReadVms
  until the Vm reports terminated: an empty list is not a state its waiter
  knows. A terminated Vm now stays readable, as it does upstream, and holds
  nothing — it is skipped by the Subnet guard and by the address count.
• The destroy then failed on the keypair, with "the keypair does not exist"
  and a gap where the id should be: the provider creates by name and destroys by
  id, and the pack read only the name.
• Four fields the provider sends that the pack declared nowhere, the worst
  being DeletionProtection — accepted and dropped, which told a client its
  machine was protected when nothing protected it. Also
  NestedVirtualization, ResultsPerPage on three reads, and ForceStop.
• feint status reported 0 in the "driven by a client" column, always.
  internal/cli declared its own shape for /_feint/conformance, with a key the
  server emits nowhere and untouched as an object where the wire carries an
  array. The decode failed, both callers fell back to empty, and the header
  comment of status.go promised exactly that number. Measured after two scw
  calls: 0 before, 2 after. The copy is deleted rather than corrected — the view
  is now one exported type both sides read.

Changed

• feint serve refuses a non-loopback address unless
  --expose-to-network says otherwise. Off loopback the anti-rebinding guard
  stops refusing anything — correctly, since it can no longer tell what is local
  — and the only output was feint dev listening on 0.0.0.0:4599. Measured: a
  cross-origin page and a forged Host both get 200 where they get 403 on
  127.0.0.1. With --vm on, what was then reachable from the network is a
  container runtime.

  This is a behaviour change a user can observe. Anyone who was deliberately
  exposing the port adds one flag; anyone who was not was exposing more than
  they knew.