[DanielThomas]

This picks up Michael Braun's work from #842, preserving his authorship, and reconciles it with the read-path invariants introduced by the preceding PRs.

When enabled, it adds snapshot iterators for list, set, and map reads. A snapshot iterator captures its holder, shard, and collection bounds once, then reads subsequent elements or buckets directly from that shard instead of resolving them for every operation:

HollowConsumer.newHollowConsumer()
        .withSnapshotCollectionIterators()
        // ...
        .build();

This also means that #345 is no longer a problem.

Benchmarks

Representative JMH measurements on JDK 25 with ZGC on this x86-64 host, traversing 2,000 lists in ordinal order per operation:

Immutable shards (WastefulRecycler)

Elements per list	Existing iterator	Snapshot iterator	Improvement
4	92.3 us	28.8 us	68.8%
16	313.0 us	50.9 us	83.7%

Recycling shards

Elements per list	Existing iterator	Snapshot iterator	Improvement
4	179.8 us	47.0 us	73.9%
16	469.7 us	117.5 us	75.0%

The set and map benchmarks also improved on this host, though the recycling cases varied more.

[DanielThomas]

List iteration, bidirectional iteration and spliteration now share the captured ordinal read logic. Splits have independent validation state. Map and set iterators capture size and bucket bounds together, and empty Java list traversal uses the JDK empty adapters.

Added coverage for cached and proxy-backed lists, refreshes, parallel resharding and empty traversal.

[Sunjeet]

• equals() and hashCode() stop agreeing. list.equals(expected) is true, but the hash codes differ. Any HashMap or HashSet keyed on such lists would miss lookups.
• spliterator() only honors the override in HollowPerfBackedList. HollowList.spliterator() falls back to super.spliterator(), which the JDK builds on iterator(), so streams skip the override too.

Suggested fix for the PR: make HollowList.iterator() use delegate.ordinalCursor(ordinal) and fall back to super.iterator() when it returns null, the same way listIterator() does. Also add the getClass() check to HollowPerfBackedList.iterator(). Streams, copies, toString() and hashCode() all go through iterator(), so that one change fixes all of them. Extending lookupSubclassRetainsItsGetBasedListIterator to cover iterator() and stream() would catch it.

[DanielThomas]

@Sunjeet Fixed both iterator paths as suggested, so they preserve custom get() behavior. I also extended the tests to cover iteration, streams, copies, toString(), equality and hash codes.

[DanielThomas]

Wired the iterator opt-in through SHARD_CURSOR_ITERATORS. Kept withShardCursorIterators() as an alias; traversal remains unchanged by default.