fix: Check Health taking 24h+ on large libraries

[plz12345]

Fixes Whisparr/Whisparr#1102

CheckHealthCommand runs for 24+ hours on a 264k scene library (~130 NFS/ZFS mounts), pinning CPU throughout, then re-triggers 6 hours later. Health checks run serially on the command executor thread, so this blocks every other scheduled task.

Note: Radarr develop is byte-identical in MountCheck.cs and DiskProviderBase.cs, so this is not Whisparr specific and there was no upstream fix to mirror. Worth upstreaming.

The bug

MountCheck calls GetMount() once per movie path, and GetMount() re-enumerated the entire OS mount table on every call, with no caching. On Linux that means reading /proc/mounts and stat'ing every mount point, so at 264k paths it did 264k full mount table enumerations, each hundreds of syscalls (and network round trips, on NFS).

The changes

Each is a separate commit and reviewable on its own.

1. Cache the mount list (the actual fix). GetAllMounts() becomes a private caching wrapper with a 15s absolute TTL over a new protected virtual FetchAllMounts(), which Mono overrides. Private so a subclass cannot accidentally bypass the cache, and so a stray override GetAllMounts fails loudly at compile time rather than silently doing nothing.

GetMount() deliberately stays per-path. MountCheck resolves mounts from movie paths specifically so the provider handles symlinks and junctions, so the number of lookups is unchanged, only their cost.

Only topology is cached. Free space is unaffected: DriveInfoMount and ProcMount read it live from the underlying DriveInfo/UnixDriveInfo on each property access.

15s is short enough that a newly mounted share still turns up in the folder browser without needing any invalidation hooks, while still collapsing a health check run from hundreds of thousands of enumerations to roughly one per 15s. This also speeds up imports, since DiskTransferService pays two lookups per file transfer.

2. Fix a RootFolderCheck N+1 (independent bug found on the way). RootFolderCheck and DiskSpaceService called GetBestRootFolderPath without passing the root folders. The result is cached per path, but on a miss it falls back to querying every root folder from the database. RootFolderCheck runs on startup, when the cache is always cold, so that was one query per movie. Fixed via the existing overload, which exists for exactly this.

3. Cheapen the per-mount scan. GetMount tested every mount with IsParentPath, which walks the path back up to the root each time. That walk only depends on the path, so it was repeated identically for all ~130 mounts. Hoisted to once per lookup.

I originally planned to also hoist PathEquals's normalization, but benchmarking showed that assumption was wrong: normalization was only ~14% of the cost, while the ancestor walk was the bulk. So PathEquals is left completely untouched, which avoids precomputing CleanFilePath per mount (it can throw, which would have changed when exceptions fire on the import hot path).

Measured over 130 mounts: ~160us -> ~62us per lookup, or ~42s -> ~16s across 264k paths.

Testing

MountCheck had no coverage at all, so the first commit adds a fixture pinning current behaviour before anything changed: the null-mount and null-MountOptions guards (MountOptions is always null on Windows, since DriveInfoMount is constructed without them), the DistinctBy on root directory, and the reported path ordering.

Both regression tests were confirmed to fail without their fix:

• mount cache: "Expected invocation on the mock once, but was 3 times"
• root folder N+1: "should never have been performed, but was 10 times" (10 movies)

New GetMount semantics tests (longest matching mount wins, exact mount root, paths with relative segments) were confirmed to pass against the old code first, so they pin existing behaviour rather than the new implementation.

I also ran a differential check of old vs new resolution across 34 real paths against this machine's 11 real mounts, including a nested pair (/System/Volumes/Data and /System/Volumes/Data/home), .. segments and trailing slashes. Identical results.

Green: Common.Test, Mono.Test, Host.Test (ContainerFixture resolves the whole DI graph, which covers the constructor plumbing), and the 144 Core.Test health check tests. Two HttpClientFixture Cloudflare tests fail, but they fail identically on eros-develop without these changes; they hit a live network endpoint and are unrelated.

Caveat

I could not reproduce a 264k library on ~130 NFS mounts locally, so the headline timing is extrapolated from benchmarks. Merging this auto-closes Whisparr/Whisparr#1102, so it may be worth having the reporter confirm the improvement on their library first.

Follow-up (not in this PR)

RemovedMovieCheck materializes 264k Movie + MovieMetadata + alt-title rows just to read one enum per row. Not an N+1, but a large allocation spike. Proper fix is a status filtered repo query.

[sonarqubecloud]

Quality Gate failed

Failed conditions
0.0% Coverage on New Code (required ≥ 80%)

See analysis details on SonarQube Cloud