You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 42e4c80
Browse filesBrowse the repository at this point in the historyBrowse files
One behaviour per path, one failure mode per mistake. All from review of
the preceding commits.
Env values are no longer path-expanded. The expansion had been added to
resolveProcessLaunch only, so a default launch and a customised one
resolved the same manifest differently -- applyLiteralLaunch never had it.
It was never needed: {models_dir} arrives absolute because the executor
expands it once when it resolves the engine's state. Dropping it leaves
the two paths identical rather than duplicating the call.
An unset models_dir is no longer published as a placeholder. It resolved
to "", so a template referencing it silently became a path under / rather
than failing. Two tests graft a bundled platform's runtime onto the fake
engine and so now need its store too; both go through one helper, since
the pair only makes sense together.
UninstallManaged checks for cancellation between engines. Its budget
bounds the commands an uninstall runs but cannot interrupt a file walk, so
the check between engines is what makes the timeout mean anything -- and
the comment claiming otherwise is corrected. The budget also rises to
match what one engine's worst case actually costs.
The install-layout test no longer reports a broken manifest as "skipped".
curl exits 22 for an HTTP error under --fail, which means the pinned URL
is wrong -- a release tag that moved -- and that is the breakage the test
exists to catch; only an unreachable network skips now. It also verifies
each archive against the manifest's sha256, so a pin left behind by a
version bump is caught here rather than by a user's failed install.
Signed-off-by: Terve <ntervalon@nvidia.com>
0 commit comments