`Path.resolve()` is a filesystem call — it lstats every component of the path.
`_resolve_dlc_path` and `safe_join` both re-resolved their ROOT on every single
call, and those run once per song, per art fetch, per scanned row.
Found while profiling a 2-fps report: on a real 50,944-song library the server
was issuing ~23,500 stat/lstat calls per second, re-walking the same three
parent directories over and over, and burning ~50% of a core doing it. It is
worst exactly where big libraries live — the library was on an NTFS-3G (FUSE)
mount, where every stat is a userspace round trip through mount.ntfs-3g (itself
visible in top). The cost was the constant re-resolution, not the work.
A root is fixed for the life of the process, so resolve it once
(safepath.resolved_root, lru_cache). Measured, 5,000 lookups against a real
library path:
before: 15,264 stat syscalls (54.2 ms)
after: 277 stat syscalls ( 0.8 ms) 55x fewer
Containment is unchanged, which is the part that matters:
- safe_join still resolves the CANDIDATE on every call — following its
symlinks IS the zip-slip / traversal defence, so it is never cached. Only
the server-owned root is.
- _resolve_dlc_path keeps its lexical containment check (deliberately does not
follow symlinks, so junction-mounted libraries keep working).
Tradeoff, documented on resolved_root: if a root's symlink is re-pointed at a
NEW target while the server runs, the old target stays in effect until restart.
Fine for a library path fixed at startup; the cache is keyed on the Path, so
switching library dir is a different key.
Tests: root resolved once across 500 lookups (the regression), a different root
is a different entry, and the containment contract re-pinned — traversal,
Windows drive-absolute, backslash, NUL, empty, and a symlink escaping the root
must still be refused. Full suite green.
