perf(cache): keep host local asset and mimetype caches in the local tier

[DeepDiver1975]

Description

⚠️ Stacked on #41732 (base is fix/ghsa-488r-findbinarypath-cache-poisoning, which introduces LocalCacheFactory and the FixedCacheFactory test double). Review/merge that one first; the base flips to master once it lands.

Moves two host-local caches out of the distributed tier and into the local one, and adds the TTLs and the invalidation that makes doing so correct.

URLGenerator::imagePath() — the cache is now resolved once in the constructor instead of calling create() on every invocation (:162), and entries get a 24 h TTL (they were unbounded). Local tiering here is strictly better than what we had: these entries are never invalidated anywhere today and the key embeds the theme name, so staleness becomes node-scoped and dies with the worker instead of persisting cluster-wide forever. Measured 0.019 ms → 0.0027 ms per call, ~60 icons per page, so the cache is well worth keeping — just not in the shared tier.

Files\Type\Loader — same tier move plus a CACHE_TTL applied at all six set() sites, which were unbounded. Worth keeping too: a 500-file listing costs 1.98 ms uncached vs 0.375 ms cached. reset() keeps clearing its own cache; that method exists to repair this process's view after a parallel-scan unique-constraint race (Files/Cache/Scanner.php:474, which clears its in-memory arrays right there), so local scope is the correct scope. Today's behaviour — one scanner losing a race nukes every node's mimetype map — is over-broad.

Latent bug fixed here because it is what makes the tiering defensible: Repair\RepairMimeTypes deletes rows from *PREFIX*mimetypes with no cache invalidation at all, so a deleted mimetype's id→name mapping survived the repair. It now takes an optional IMimeTypeLoader (defaulted, so third-party construction keeps working) and calls ->reset() at the end of run() — but only when something was actually repaired, so a no-op upgrade doesn't throw every node's map away. Single construction site lib/private/Repair.php:129 updated.

core/Command/Upgrade.php — the blanket create()->clear() only ever reached the distributed tier. It now clears both, and the stale TODO is rewritten. Residual limit stated plainly in the code comment: clearing local from CLI only clears the node running occ, so bounded staleness now comes from the new TTLs plus worker restarts. That is exactly why the TTLs are in this PR and not a follow-up.

Untouched on purpose: the locking cache (2.96 ms DB vs 0.023 ms memcache per file — 130×) and everything genuinely cluster-shared.

Related Issue

• Fixes n/a — follow-up hardening from the GHSA-488r-cjpq-p3vc investigation

Motivation and Context

Host-local values in a network-reachable, protocol-injectable backend are a standing supply of write primitives like the one in #41732. Neither of these two needs to be shared between nodes, and moving them also makes their staleness bounded rather than permanent.

How Has This Been Tested?

• test environment: PHP 8.3.32, sqlite, plus a run with memcache.local => APCu + memcache.distributed => Redis — CI sets no memcache.* and is blind to tiering. (Note for anyone reproducing: OC\Memcache\APCu::isAvailable() gates on the ini key apc.enable_cli, not apcu.enable_cli, or the local tier silently degrades to NullCache on CLI.)
• test case 1: tests/lib/Files/Type/LoaderTest.php — the old ->expects($this->once())->method('create') mock is replaced with FixedCacheFactory; new testUsesTheLocalCacheTier() (asserts createLocal('mimetypes') once, create/createDistributed never) and testCachedEntriesExpire() (asserts set() receives Loader::CACHE_TTL).
• test case 2: tests/lib/UrlGeneratorTest.php — new testImagePathIsCached() (asserts the resolved path, that the theme name is part of the key, and that a planted value is served on the second call), testImagePathsUseTheLocalCacheTier(), testImagePathThrowsForMissingImage().
• test case 3: tests/lib/Repair/RepairMimeTypesTest.php — testMimeTypeCacheIsResetAfterRepair() (version 8.0.0.0 → reset() once) and testMimeTypeCacheIsKeptWhenNothingIsRepaired() (version 10.0.0.0 → never).
• test case 4: negative verification — each new test was re-run with the source change temporarily reverted and confirmed to fail (3 failures), then restored. A test that passes either way proves nothing.
• test case 5: real-instance A/B — after the change imagePath and mimetypes keys exist only in APCu, with redis KEYS returning 0 for both namespaces; icons and mimetypes still resolve correctly.
• test case 6: full tests/lib — 7149 tests, 40221 assertions, no new failures. make test-php-style and phan clean.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Database schema changes (next release will require increase of minor version instead of patch)
• Breaking change (fix or feature that would cause existing functionality to change)
• Technical debt
• Tests only (no source changes)

Checklist:

• Code changes
• Unit tests added
• Acceptance tests added
• Documentation ticket raised: n/a
• Changelog item, see TEMPLATE

🤖 Generated with Claude Code

[phil-davis]

Looks reasonable, good for these cached things to stay in a cache only local to the node.