Add local dev folder to POCv2 - #19

Code comments cited files that are gitignored, so they were dead pointers for
anyone reading the repository. They now cite moby/dockerode behaviour in its
own terms or the committed requirements sections instead. The reasoning is
unchanged; only the references are.

Probe image resolution moves onto a shared resolveTaggedBuild helper, also now
used by the run-start path it was duplicating. This drops an arbitrary-tag
fallback: an Actor whose only build is not tagged latest is refused at
registration, matching what a tag-less run already does rather than succeeding
against a build no run could reach.

Adds coverage for the build-time working-directory persist, both when the
build reports one and when it does not and a stored value must survive.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

A tag whose build record was deleted reported "Actor has no build tagged X",
which was wrong: the tag exists. Build resolution now reports the two cases
apart, both still 404 record-not-found, and both are covered by tests.

The default build tag becomes one exported constant that the run route and the
run service import, replacing three hand-synced 'latest' literals.

Adds a services-layer test that a run's mount is derived from the Actor's
stored fields, covering both set and cleared.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

…edAt

Registering a dev folder bumped the Actor's modifiedAt, which is exposed on
/v2 -- the emulated API could observe a purely local registration. Dev-folder
writes no longer touch it, and clearing an unregistered Actor writes nothing.

The dev-folder pipeline moves out of the actors service into its own module,
which lets it reuse the builds service instead of duplicating a registry read
to avoid a cycle. HTTP statuses and error types move to the route layer.

A tag whose build was deleted reports the same bare not-found the run route
gave before this branch, so run-start behaviour is unchanged throughout.

Probe containers are labelled and swept by orphan reconciliation, the driver's
probe method is required rather than optional, whitespace-only paths are
rejected instead of silently clearing, console styles move into the shared
template, and the e2e test copies the sample Actor instead of editing it.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

Sweeping probe containers alongside run containers put both label keys in one
listContainers filter, but the daemon requires a container to carry every key
given in a single call, so the query matched nothing and reconciliation stopped
reclaiming orphaned run containers -- and their anonymous volumes with them.
Each label key now gets its own query and the results are unioned by id.

The stub in the unit test now models that daemon-side behaviour, so a combined
call returns nothing and fails, and a further test pins one value per call.

The dev-folder route module regains the mount(router, deps) shape its siblings
use, with the server owning the /actor-runtime prefix as it owns /v2.

Also corrects the rejection message for an Actor whose build is not tagged
latest, and the requirements and comments that repeated its framing.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

The test installed dependencies into the folder it then bind-mounted, so the
Actor resolved its imports from the host copy and the run would have passed
whether or not the image's node_modules survived the mount. The copied folder
now has node_modules removed before registration, so resolution can only come
through the anonymous volume.

Types the dockerode stubs in the driver unit test, which had drifted to casting
empty mock-call tuples; the build does not cover test files, so this went
unnoticed.

Documents the probe label and its startup sweep, records that host-level volume
and container invariants may be read from the Docker CLI since no Apify command
observes them, and corrects the claim that the runtime is always containerized.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

startRuntimeContainer binds fixed host ports, and vitest runs test files in
parallel by default, so the two e2e files raced for 3333/3000 and whichever
lost died on "port is already allocated". test:e2e now disables file
parallelism, and the port-binding site says why so a third file does not
rediscover it. The unit and integration suites keep running in parallel.

Also corrects the driver doc that still claimed one bind-mount site, drops a
cast the stub's own parameter type makes unnecessary, states the registration
precondition in the README as the latest tag rather than any build, and rewraps
a CLAUDE.md bullet that split an inline code span.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

Docker accepts a regular file as a bind source, so a file path passed the
dev-folder check and only failed later as a broken run. The probe now binds
<path>/. , which the daemon rejects for a file while still reporting a missing
path as missing, and the route answers with a message naming the real cause.

The image working directory moves from the Actor to the build that produced it,
and a run mounts against the working directory of the build it resolved. An
Actor with several tags could otherwise mount at a path taken from a different
tag's image while the console reported the mount as applied.

The endpoint is also served under /v2 so it can be reached without a ../ prefix,
since apify api builds its URLs onto a /v2 base. Both mounts share one handler
and one auth pass; the documented form drops the dots.

Adds an end-to-end case for the workflow this exists for: two runs of the same
build, recompiled in between, with no push and no rebuild.

Formats CLAUDE.md so format:check passes, and forces the end-to-end re-pushes,
which apify push refuses once a build has bumped the Actor past the pushed
files' timestamps.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

The bind-mount sections described how today's code works -- library calls,
Docker API field names, the daemon's own error strings, internal file names --
rather than what is required. They now state outcomes: the path must be
absolute, must exist on the host, and must be a directory; being unable to
verify is never reported as missing; the mount applies only when the folder and
the resolved build's working directory are both known; a folder that has gone
away fails the run visibly; the image's dependencies survive the mount.

The constraint that the runtime cannot judge a host path from its own
filesystem stays, without prescribing the workaround.

Restores the note that the console is no longer strictly view-only, which the
previous trim dropped while the console does write.

Comments and test titles quoting the old wording are repointed. Formats api.md
so format:check passes.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv

Registration needed an image only to create the throwaway container that asks
the daemon whether the submitted path is a mountable directory. Borrowing the
Actor's latest build for that made registration depend on a prior build and
turned a folder problem into a message about build tags, while the image itself
contributes nothing to the check.

The driver now builds and owns a minimal image for that purpose, so a folder can
be registered against an Actor that has never been built.

The reported status is now just the registered folder. Whether a mount applies
depends on which build a run resolves, so an Actor-level field could not answer
it honestly.

The tag-less run default is untouched -- that belongs to the run route, not to
this feature.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011L4VcFqN9UZUbVRMvQSugv