[DanielDarrenJones]

While indexes are built, CollectionEntriesStore::makeItemFromFile() keeps every entry it makes in Blink's structure-entries (so URI building can find entries that aren't in the Stache yet). getCachedItem() then unserialises a second, full copy of each entry from the cache. On a cold warm, a terms index that loads every collection's entries ends up holding all of them twice.

This changes getCachedItem() so that, only while an index is loading (Index::isLoading()), it hands out a shallow clone of the entry already in memory instead of unserialising another copy. The clone has the same per-object caches cleared that Entry::__sleep() leaves out of the cached copy (cachedKeys, computedCallbackCache, siteCache, augmentationReferenceKey), so callers get the same object they would have got from the cache, and its data arrays are only copied if they change it. Outside index loading nothing changes.

Cold Stache::warm() on 6.40.0, using a production site's content (4,111 entries across 11 sites, 113 MB of YAML; median of 3 runs, Apple M1 Pro, PHP 8.4, file cache store):

	time	peak memory
6.40.0	49.5 s	1,710 MB
this PR	41.3 s (−17%)	841 MB (−51%)

%%{init: {"xyChart": {"width": 700, "height": 220}}}%%
xychart-beta horizontal
    title "Peak memory, cold Stache warm (MB)"
    x-axis ["6.40.0", "This PR"]
    y-axis "MB" 0 --> 1800
    bar [1710, 841]

Loading

%%{init: {"xyChart": {"width": 700, "height": 220}}}%%
xychart-beta horizontal
    title "Time, cold Stache warm (seconds)"
    x-axis ["6.40.0", "This PR"]
    y-axis "Seconds" 0 --> 55
    bar [49.5, 41.3]

Loading

Every entry's uri() and url() are identical with and without the change (all 4,111). Handing out the clone at all times (not just while loading) gets to 754 MB, but limiting it to index loading keeps normal requests exactly as they are.

How this differs from #3984

#3984 (3.4, closed in 2024 without being merged) also used Blink and clones, but for a different problem:

• Goal. Store/Get items in/from Blink instead of always calling from Cache. #3984 cut repeated cache reads on every request; this cuts memory while the Stache is built.
• Scope. Store/Get items in/from Blink instead of always calling from Cache. #3984 changed BasicStore, so every store, on every request. This only touches CollectionEntriesStore, and only while an index is loading.
• Mechanism. Store/Get items in/from Blink instead of always calling from Cache. #3984 added a new in-request cache: every item read or written also went into Blink. This adds nothing new; it reuses the copy that makeItemFromFile() already keeps in structure-entries, instead of unserialising a second one.
• Copies. Store/Get items in/from Blink instead of always calling from Cache. #3984 returned a plain clone, which keeps the object's memoised computed values and site cache. This clears exactly what Entry::__sleep() drops, so the clone behaves like a copy read from the cache.
• Side effects. Because object identity changed everywhere, Store/Get items in/from Blink instead of always calling from Cache. #3984 also had to flush Blink before queue jobs, clear Blink on taxonomy save and switch about ten tests from assertSame to assertEquals. Here, normal requests are unchanged and one strict Blink mock needed to allow the extra get().

On a cache miss, BasicStore::getItem() already returns the very instance it put in Blink, so sharing in-memory entries isn't new; this change hands out clones rather than sharing.

Tests: three new tests in EntriesStoreTest (two fail without the change; the third checks that reads outside index loading still come from the cache). EntryTest::it_clears_blink_caches_when_saving mocks Blink strictly, so it now also allows the get() that indexes loaded during the save make; it still requires the entry to be forgotten and put again. The full 6.x suite passes (11,464 passed, 12 skipped, as on 6.x).