perf(covers): stop enumerating MediaCover/Books for every book without a cover folder

[jordanfelle]

Problem

GET /api/v1/book?authorId=N (what the book and author pages load) is quadratic in the size of the catalogue. On a live instance the time per book is flat at about 3 ms for small authors, then falls off a cliff:

Books for the author	Time
363	1.0 s
1,059	2.7 s
2,522	no response in 100 s
12,269	no response in 280 s

The database is not the bottleneck. pg_stat_statements showed no activity for 12 s or more mid-request, and the controller's own mapping is linear (a benchmark of the real GetBooks and MapToResource with stubbed services maps 5,000 books in about 110 ms).

Cause

Stack snapshots taken while a request was running put every request thread in MediaCoverService.TryReadBookCoverMetadata -> DiskProviderBase.FileExists -> ResolveExistingFilePath -> ResolvePathSegment. TryReadBookCoverMetadata looks up cover-metadata.json per book. When the file does not exist, FileExists falls back to a case-insensitive path resolution that lists and normalises every entry of the parent directory. Here that is MediaCover/Books with 194,613 folders (one per book), and most books have no cover folder: 2,286 of the 2,522 books of the author above. That is about 445 million comparisons for one request.

Fix

Check FolderExists (a plain Directory.Exists) for the book's cover folder first, and cache the negative result the same way the missing-file case already was. Books that do have a cover folder behave as before.

Verification

• New regression test converting_a_book_without_a_cover_folder_should_not_probe_for_its_metadata_file: fails without the change, passes with it.
• All 66 MediaCover tests pass.
  Measured live after deploying: GET /api/v1/book?authorId=552 (2,522 books) went from more than 100 s (timeout) to 1.36 s, and authorId=2698 (12,000 books) from never returning to 5.9 s.

Review update: the read path now writes the negative cache entry with TryAdd, so a cover download that stores its metadata between the disk check and the cache write is no longer overwritten by a stale 'missing' result (test: a_concurrent_cover_write_should_not_be_overwritten_by_the_missing_folder_negative_cache_entry, deterministic via an OnFolderExists hook on the test disk stub) (3,022 core tests pass).

[jordanfelle]

Live timings with this change deployed (library of 346k books, 195k folders in MediaCover/Books), one request each:

Request	Before	After
GET /api/v1/book?authorId=552 (2,522 books)	timed out past 100 s	1.36 s (1.28 s warm)
GET /api/v1/book?authorId=2698 (12,000 books)	never returned	5.9 s
GET /api/v1/book/402436	-	0.06 s