Fix #136: audit fan-out + legacy migration subdir + cache_path warnings

[Brandon-Haney]

Fixes #136.

Summary

Three commits addressing the root cause and two contributing configuration issues behind the 26-minute audit cycles reported in #136. The reporter's server (1.23M cache files) was issuing ~4.9M os.path.exists probes per audit, saturating the 5-minute WebCacheService refresh thread and keeping Unraid array disks spun up indefinitely.

1. Audit fan-out fix (perf) — primary fix

MaintenanceService.run_full_audit() previously issued 3-4 os.path.exists probes per cache file:

• _check_plexcached_backup() — one probe on /mnt/user0/...plexcached
• _check_array_duplicate() — one probe on /mnt/user0/...
• A second duplicates-only pass repeating _check_array_duplicate() on every cache file
• _get_orphaned_plexcached() separately walked every array dir

On cold Unraid arrays each probe blocks on disk spinup; even on a warm array the aggregate wall time on a 1.23M-file library was ~26 minutes per audit.

Fix: walk the array disks exactly once inside _get_orphaned_plexcached, build array_files_set and plexcached_set during that same walk, and answer all backup/duplicate questions via O(1) set lookups. The duplicates pass is collapsed into the main cache-file loop so cache files are iterated once instead of twice. Zero extra array I/O vs. the walk that was already happening.

Also adds observability so this regression can't silently creep back:

• run_full_audit logs total wall time at INFO and per-phase timings (scan / array_walk / main_loop) at DEBUG.
• WebCacheService.refresh_all() logs a WARNING when a single refresh cycle exceeds REFRESH_INTERVAL_SECONDS, surfacing the exact symptom from Dashboard Stuck Loading - Unraid Docker Version #136.

2. Legacy→path_mappings migration subdir bug

The reporter's migrated config had:

{
  "real_path": "/mnt/user/Media/",
  "cache_path": "/mnt/cache/"        // ← should be /mnt/cache/Media/
}

Because cache_path pointed at the cache drive root, the audit walked the entire SSD (appdata, docker, system, every container) — 1.23M files. The migration now mirrors the real_path subdirectory into cache_path. Ships with a read-only detect_path_mapping_health_issues check and dashboard banner so existing bad configs get flagged loudly on startup without auto-rewriting user state.

3. cache_path validation (warn, don't block)

PathMapping.cache_path values like /mnt/cache (bare root) or /mnt/user/... (FUSE) cause audit fan-out or make _check_array_duplicate silently wrong. The final approach is warn-only, never blocking — some configs legitimately use /mnt/cache/ as a flat root or /mnt/user/... in containers where /mnt/cache isn't mounted.

• New SettingsService.warn_cache_path() helper returns a human-readable warning for risky values, or None otherwise. Exempts /mnt/user0/ and custom mount points.
• Save endpoints (POST /settings/paths, PUT /settings/paths/{index}) call the warner, log a WARNING, and always persist the user's value.
• auto_fill_mapping() refuses to trust settings['cache_dir'] when it starts with /mnt/user/ — falls back to /mnt/cache to prevent stale legacy values from propagating into newly auto-generated mappings.
• Dashboard health banner (shared with Fix hardcoded config path to use script location #2) persistently surfaces the warning until the user fixes it or decides to keep the value.

Commits

• Fix legacy migration cache_path subdir and add health detector (#136)
• Warn (non-blocking) on risky cache_path values and tighten auto-fill (#136)
• Fix audit fan-out: single array walk with O(1) set lookups (#136)

Expected impact

• Reporter's 1.23M-file library: ~26 min → a few seconds per audit cycle (single O(n) walk instead of ~4.9M probes).
• Dashboard / maintenance tabs stop stalling behind the background refresh thread.
• Unraid array disks no longer stay spun up by the refresh loop.

Test plan

• New unit tests for the legacy migration (tests/test_config_migration.py, 12 tests) covering the exact real_path=/mnt/user/Media/ → cache_path=/mnt/cache/Media/ shape.
• New unit tests for cache_path warnings + auto_fill_mapping tightening (tests/test_libraries_settings.py, 28 tests).
• New end-to-end TestRunFullAuditFanOut tests (tests/test_maintenance_service.py): one verifies unprotected/duplicate/orphaned detection is preserved against real tmp_path files, the other patches os.path.exists and asserts it is not called per individual cache file — guards against reintroducing the fan-out.
• Existing _check_plexcached_backup / _check_array_duplicate helper tests still pass (helpers are untouched; only the audit loop bypasses them with set lookups).
• Full suite: 88/88 maintenance + migration + libraries tests pass on the PR branch.
• Verification on the reporter's 1.23M-file production library — compare os.path.exists call count and wall time before/after (reporter willing).

Notes for reviewer

• _get_orphaned_plexcached() signature changed to return a 4-tuple (orphaned, extensionless, array_files_set, plexcached_set) and accepts an optional cache_files arg so run_full_audit doesn't re-scan. All five in-repo callers updated.
• The cache_path warnings are intentionally non-blocking per team discussion — blocking the save path would break legitimate flat-root and container configurations.
• No API changes for callers of run_full_audit().

[StudioNirin]

Can't see any issues, I'll leave this for you to merge in case there was anything you wanted to double-check