Map named volumes to native container volumes (fixes #96)

[kylebrowning]

Problem

The named-volume branch of composeVolumeToRunArgs bind-mounted a host directory and mounted it at the parent of the declared destination via destination.deletingLastPathComponent(). Two failure modes fall out of this:

1. Mounts over / (issue volumes mounts parent directory than we expect, may tries to mount root directory #96). A single-segment destination collapses to /. So redisdata:/data mounts the volume over the container's entire root filesystem, shadowing the image's entrypoint. Redis fails to start with:

   failed to find target executable docker-entrypoint.sh
   

2. No chown on virtiofs. Even for nested destinations, the host bind mount uses virtiofs, which the guest cannot chown. Images whose entrypoint chowns its data directory (postgres, redis, …) die with:

   chown: .: Operation not permitted
   

A side effect of (1) for nested paths (e.g. pgdata:/var/lib/postgresql/data) is that data was being mounted at /var/lib/postgresql — the parent — so it wasn't actually persisted at the path the user asked for.

Fix

container 1.0.0 supports named volumes natively and auto-creates them on first run. This maps a Compose named volume straight to a container named volume named <project>_<source> (mirroring Docker Compose's namespacing), mounted at the declared destination.

Native volumes are block devices the guest owns, so chown works, and the destination is no longer rewritten to its parent. This also retires the now-outdated warning text ("the container tool does not support named volume references"), which predates 1.0.0.

Verification

Against a real docker-compose.yml with a redis service using redisdata:/data:

• Before: container crashes on startup — failed to find target executable docker-entrypoint.sh.
• After: backend_redisdata volume is auto-created and redis reaches Ready to accept connections / redis-cli ping → PONG.

Tests

Adds coverage to ComposeVolumeTests (the named-volume branch previously had none):

• maps to a native, project-namespaced volume
• volumes mounts parent directory than we expect, may tries to mount root directory #96 regression: a single-segment destination must not collapse to /
• nested destination + mode is preserved
• missing project name is skipped

All 137 static tests pass. (The .containerDependent dynamic suites fail in my environment with a pre-existing No such file or directory at setup — reproduced identically on a clean main, unrelated to this change.)

Note / out of scope

A postgres service using a named volume will still hit the postgres image's initdb: directory ... not empty ... lost+found guard, because container's native volumes are formatted filesystems with a lost+found. That's image/platform-specific and is worked around user-side by pointing PGDATA at a subdirectory of the mount; it's independent of this fix.

[Cyb3rDudu]

Nice - this is a real fix. The destination.deletingLastPathComponent() collapsing /data down to / and shadowing the entrypoint is exactly #96, and moving to native volumes + the declared destination is the right call. The tests covering the regression and the nil-project case are great too.

Heads-up on an overlap: #119 rewrites this same branch with a more complete take — it honors the top-level volumes: block (name overrides, external: true skip-create, driver / driver_opts / labels) and replaces the createVolumeHardLink loop with a native container volume create. As written here those get dropped, and createVolumeHardLink in ComposeUp would still run, creating host dirs that nothing mounts anymore.

What this PR gets right that #119 doesn't: the <project>_<source> namespacing. #119 names the volume just data, so two projects would collide.

My plan is to take #119's handling as the base and fold your <project>_<source> default into it (I've left that suggestion on #119) - one coherent design instead of two. So I'm going to hold this one rather than merge it as-is, but the namespacing idea is yours and I'll credit it when I wire it in. Really appreciate you digging into #96; happy to ping you when the combined version is up so you can sanity-check it.

[Cyb3rDudu]

closing — #119 merged with your _ namespacing folded in, so this is covered. thanks for the idea!