feat: import existing files on mount + fix path-as-tag

[Dhravya]

Summary

• Import existing files on mount: When mounting to a non-empty directory, detects pre-existing files and prompts the user to import them into the Supermemory container. Adds --import-existing flag to skip the prompt. Files are scanned before the mount (which shadows them), then imported after the initial pull so server-side files take precedence.
• Fix smfs mount .: The positional arg is now detected as a path when it looks like one (., .., contains /, or is an existing directory). The container tag is derived from the directory name and the path is used as the mount point.
• Fix smfs unmount .: Canonicalizes the path before walking up to find the .smfs marker, so relative paths like . resolve correctly.

Test plan

• smfs mount mytag — works as before (tag = mytag, path = ./mytag/)
• smfs mount . — mounts current directory, derives tag from dir name
• smfs mount ./somedir — mounts somedir, derives tag from somedir
• smfs mount mytag --path /some/dir — explicit path still works
• Mount to non-empty dir in TTY — prompts for import
• Mount to non-empty dir in non-TTY — prints warning
• --import-existing skips prompt and imports
• smfs unmount . — finds .smfs marker and unmounts correctly
• cargo test — all 143 tests pass

🤖 Generated with Claude Code

[vorflux]

Reviewed — found 3 medium-severity issues.

---

Review with Vorflux

[vorflux]

looks_like_path misclassifies existing directory names as path-as-tag, breaking --path disambiguation.

looks_like_path returns true for any name that happens to match an existing local directory (std::path::Path::new(s).is_dir()). When that triggers, resolve_tag_and_path immediately errors if --path is also provided:

if explicit_path.is_some() {
    anyhow::bail!("cannot use both a path as the tag and --path");
}

This means smfs mount docs --path /tmp/mnt fails whenever a ./docs directory already exists in the working directory, even though docs is the intended container tag and --path was explicitly supplied to disambiguate. The explicit --path should take priority over the filesystem probe — when --path is present, skip the is_dir() check entirely (or treat the positional as a tag unconditionally).

[vorflux]

collect_files_recursive reads entire file contents into memory before any size gate is applied — can cause large memory spikes or OOM on startup.

std::fs::read(&path) buffers the full file into a Vec<u8> for every file in the tree. The 100 MB cap only fires later inside SqliteFile::write (called from import_file). So a directory with a single 2 GB file — or many large files — will exhaust memory at startup before any rejection happens.

Fix: check entry.metadata()?.len() before calling std::fs::read, and skip (with a warning) any file that exceeds the size limit. Alternatively, stream the content in chunks rather than buffering the whole tree upfront.

[vorflux]

--import-existing combined with --no-sync can silently duplicate remote documents.

The comment on line 158 says "Import after initial pull so we skip files already in the container", but when cfg.no_sync is true there is no initial pull. On a first mount or clean cache, any local file that already mirrors a remote document will have no remote_id after import, and the push worker will POST it as a new document instead of skipping or PATCHing the existing one — creating duplicates in the container.

Either reject the combination at startup:

if cfg.import_existing && cfg.no_sync {
    anyhow::bail!("--import-existing cannot be used with --no-sync: \
                   duplicate detection requires an initial pull");
}

or perform a targeted remote existence check before importing each file when no_sync is set.