Compute provider qualification

A provider in the catalog is an implemented adapter, not proof that an installed release can rent and recover a machine in your account. Settings, Skarbiec credentials, provider admission, quota, and a release-scoped live run are distinct checks.

stado capabilities --json lists implemented, partial, external, and planned provider facets; it is a catalog read, not a credential check or live acceptance. stado config validate checks admission and required settings, not whether a vendor currently grants capacity.

From the Stado checkout, node tests/providers/gpu_cloud.mjs exercises real stado config validate refusals for every GPU/cloud vendor and a read-only stado resources show --provider <vendor> --json against one selected vendor's actual API. Unlike instances list, this inventory does not collect expired capacity publications while reading. Set STADO_BIN to the exact executable under test, STADO_GPU_TEST_PROVIDER to that vendor, and select a STADO_CONFIG profile which enables it and has its required settings. Its account credential must be available through the cloud-<vendor> role in Skarbiec. Run it separately for each vendor; no provider is inferred from the machine's environment.

node tests/providers/gpu_cloud.mjs

The runner refuses absent test inputs, an unknown vendor, a dirty checkout, or an executable whose embedded source revision differs from the checkout. It records the source revision, binary and runner hashes, commands, statuses, provider responses and outcome under the checkout's ignored .build/ directory. A failed live read is a failure, not an empty fleet. Without a selected live vendor, the report is blocked and exits nonzero; configuration refusals alone never earn a passing integration verdict.

This read-only run does not qualify provisioning, agent startup, job execution, reaping, or deletion. Those need a separate real isolated lifecycle run and final external-state inspection before an adapter can be called operational. It neither rents a machine nor spends on inference. Stado Desktop shows provider admissions, instances and quota through the same native operations and keeps the refusal receipt; it is not evidence of a successful cloud lifecycle on its own. See Quota diagnostics for an unreadable GCP accelerator census.

For GCE's single-instance lifecycle read, only a real HTTP 404 means the VM is absent. A successful response without a string status is refused with the instance, zone, path and observed field type. An unfamiliar status is not translated to “absent” when deciding whether an instance is running; the refusal names that status so recovery does not silently assume deletion.

If GCE accepts an insert but the operation wait fails, Stado reads that specific instance before trying another zone. A booting or running VM returns its existing reference; HTTP 404 permits the next zone. A present VM in another state or an unreadable existence probe refuses further placement, naming both the insert and probe outcomes. It does not risk renting a second VM for the same job on an unknown first result.