sky-monitor and rvlite tests fail only in a workspace build: ruvector-core/storage is unified on

[ohdearquant]

Sixteen tests in Tests (core-and-rest) fail only when the shard is run as one cargo
invocation, and pass when their crate is run alone:

cargo nextest run -p sky-monitor      29 passed
cargo nextest run -p rvlite           71 passed, 2 skipped

versus, in a run of the whole shard's package set:

sky-monitor  9 failed   Vector(DatabaseError("Database already open. Cannot acquire lock."))
rvlite       7 failed   RvLiteError { message: "Failed to create vector database:
                        Database error: I/O error: No such file or directory (os error 2)" }

Cause

Both crates depend on ruvector-core with default-features = false and a minimal
feature set:

• examples/sky-monitor/Cargo.toml:24 — features = ["simd", "parallel"]
• crates/rvlite/Cargo.toml:15 — features = ["memory-only"]

and both pass a placeholder storage_path, with comments saying it is ignored:

• examples/sky-monitor/src/indexer.rs:53-55 — "sky-monitor-tracks.mem", commented
  "Ignored by the in-memory backend (ruvector-core is built here without the storage
  feature); kept for API completeness."
• crates/rvlite/src/lib.rs:171 and crates/rvlite/src/sql/executor.rs:135 —
  "memory://".

Cargo features are additive and unify per build. ruvector-core's default set includes
storage, and other members of this workspace depend on it with defaults, so in a
single-invocation workspace build ruvector-core/storage is on for everybody.
VectorDB::new then takes the #[cfg(feature = "storage")] path unconditionally
(crates/ruvector-core/src/vector_db.rs:38-73) and opens the placeholder as a real
path. "memory://" is not a directory, hence rvlite's ENOENT;
"sky-monitor-tracks.mem" is a single fixed relative path shared by every test
process, hence sky-monitor's lock contention under a per-test-process runner.

So default-features = false does not do what both call sites assume, and the failure
appears only in the exact configuration CI uses.

Why it is invisible today

Tests (core-and-rest) is cancelled at timeout-minutes: 240 while still compiling,
so none of these results are reachable. #784, #786 and #787 are what let that job get
to the test phase; these sixteen failures are waiting behind them.

Suggested direction

The choice between in-memory and persistent storage is a property of the caller's
request, not of who else is in the build graph, so it probably belongs in DbOptions
rather than in a cargo feature — for example honouring a memory:// path (or an
explicit in-memory option) even when storage is compiled in. memory-only already
exists as a feature name in ruvector-core/Cargo.toml but carries no code, which
suggests that intent was already there.

A narrower stopgap is to give each VectorDB a unique writable path, which fixes
sky-monitor cleanly. It is a poorer fit for rvlite, where "memory://" is library code
and writing files would be a behaviour change rather than a test fix.

Happy to send patches for whichever direction you prefer.