Fixed
-
The address a client reads now answers, on every provider and in both
runtime modes (#116, #117). A server carrying a flexible IP published one
address while the machine held another, or none: the routing was recorded
before the machine existed and never replayed at poweron, and a machine with
no private NIC booted on the operator's own profile bridge, where applying a
firewall rule re-plugged the interface and cost it the DHCP lease. Sossh
to the published address timed out while the container was up and answering
its neighbours.The address a client reads stays the provider's own — TEST-NET-3, -2 and -1,
one block per pack so two emulated clouds on one host can never route the
same /32 — and it is now genuinely routed, from the first boot, on a network
the emulator owns. Publishing the runtime's DHCP address was rejected: it
desynchronisesserver.public_ips[].addressfromGET /ips/{id}, which the
real cloud never does, and it varies with--vmand with the host.dynamic_ip_requiredwas decoded, echoed back in the response and read by
nobody; it allocates an ephemeral address now, released at stop. Because the
field was read, it never appeared inunread_request_fields— the one
blind spot the contract gate has. -
The login is proven on all three packs, not asserted on two.
ssh.sh
existed for Scaleway alone, sooutscaleand Exoscale's template
default-userwere names in a table nobody drove. There are three suites
now, each registering its key through its own provider's API and opening a
real session, and each fails rather than skips when a runtime is present —
the Scaleway one used to skip there, which is how a broken address shipped
under a green run. -
What OVN cannot do is declared rather than written as a limit. A
subnet-internal address answers the host in bridge mode and not in OVN,
because the router that separates two VPCs also SNATs the host's connections
on the way back — measured, with sshd up and answering its neighbours while
the host read the port as closed.capabilities.private_from_hostsays
which, so a probe skips with a reason instead of failing, and the routed
public plane crosses in both modes, which is why every pack's ssh chain uses
it. -
A whole Scaleway product answered
404 page not foundin plain text.
Reported by @vde-dis on #74, measured under a real OpenTofu apply:
/lb/v1/…and/vpc-gw/v2/…fell through to net/http's default mux, and the
Scaleway SDK reads the content type first and drops a body that is not JSON,
so the caller got404 Not Foundand nothing else. Every other refusal
measured answers in the provider's own dialect.The prefix list declared only the five products this pack serves, while
Scaleway's URL space has sixty-two. It now declares the space rather than the
served part, extracted from the SDK's generated request paths — not its
directory names, which differ: the SDK saysvpcgw, the URL says/vpc-gw/.The guard that was supposed to prevent this could not see it.
TestEveryRouteFallsUnderADeclaredPrefixwalks the routes the pack mounts, so
a product with no routes is invisible to it by construction, and its
comment claimed it stopped "a whole product answering net/http's plain text".
Falsified: with the list back to five, the new test fails and that one still
passes, which is the blind spot demonstrated rather than argued. -
An unknown image identifier no longer boots a substitute (#83). With a
runtime configured, all three packs silently replaced an image no catalogue
held — ask for Alpine, boot Ubuntu — while the API kept reporting the
identifier the client sent; Scaleway's resolution matched labels by
substring, socentos,rockyandubuntu_focalall became Ubuntu 22.04.
The create still succeeds, asdocs/limits.mdpromises for hardcoded
production identifiers, but the boot now refuses: the machine reaches the
provider's own failed state and the log names the identifier. Image and
login now resolve together — root on Scaleway,outscaleon Outscale, the
template'sdefault-useron Exoscale — and the Scaleway marketplace answers
one fixed UUID per label, so a label resolved through it by Terraform still
names the distribution it chose. An image the client registered through
Outscale'sCreateImagerefuses the same way — the emulator serves its
record and holds no disk contents — and the log says which of the two cases
it is.
Security
-
Two CI installs took whatever was newest.
pipx install uvin the weekly
drift workflow was pinned to nothing at all — in the job that regenerates
committed artefacts and opens a pull request with them — and commitizen was
pinned by version rather than by hash. Both install from a requirements file
with hashes now, andTestTheWorkflowsPinTheSameToolsAsMisefails when a
version there stops matching the file that owns it:mise.tomlfor uv, the
pre-commit hook for commitizen, so a workstation and a runner cannot run
different tools. Reported by OpenSSF Scorecard's Pinned-Dependencies. -
SECURITY.mdcarries a reachable address. It described how to report and
linked nothing, so a reader had to know where GitHub keeps the form. It names
the advisory URL now, and a second route for anyone GitHub is unavailable to.