MB-72489: top-N collector optimizations - #2385

…ion)

collectStoreHeap.Final was 11.6% of total CPU when profiling k=1000 BM25
queries. Extracting k docs via repeated removeLast (heapsort) requires
O(k log3 k) comparisons with scattered pointer dereferences at each heap
level - poor cache behavior. Fix: sort.Slice in-place (pdqsort) then copy
heap[skip..skip+size-1] sequentially. Final cum cost: 1.29s -> 0.52s (-60%).

Note: the original perf-gar commit (89fa0a1 on perf-gar-v17-only) also
specialized the score-descending comparator; that half is already on master
via MB-72489 (#2381) as getOptimalCollectorCompare, so this cherry-pick
carries only the Final() change.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
(cherry picked from commit 89fa0a1)

Adds 9 unit tests for the collectStoreList type in search/collector that
had 0% coverage before this commit. Tests cover round-trip insertion order,
size-capped eviction (AddNotExceedingSize), skip-based pagination (Final),
Internal() ascending traversal, removeLast worst-doc eviction, single-element
and equal-score edge cases, and fixup error propagation — all using existing
scoreDesc / makeScoreDoc helpers from heap_test.go.

These tests have no dependency on any perf-gar change and can be rebased
or cherry-picked independently.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
(cherry picked from commit 7de1081)

Reset saves five backing arrays and a map, wipes the struct with
*dm = DocumentMatch{}, then restores them. That wipe is a ~240-byte duffzero, and
most of what it clears is either about to be overwritten or already zero.

Zero the nine fields that are NOT restored below instead. Same postconditions,
no duffzero.

Reset runs once per collected document, so this is on the hottest per-document
path there is — it shows up on single-term queries, not just disjunctions.

Split out of perf-gar-mod-fix-2's 5d9689f, which bundled this with a §15
minSegCeiling guard. The two are unrelated: the guard needs the per-segment score
ceilings and therefore the v18 MaxTFNorm sidecar, while this is pure in-memory work
that applies to any segment format. Only this half belongs on a v17-only branch.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Guard pointer/string/map field zeroing in Reset with nil/empty checks so the
common case (no explain, no fragments, no field highlights) skips nil-to-nil
stores and their GC write barriers (~5 cycles per field, across the millions
of Reset calls in a large result collection).

Note: the original perf-gar commit (1b30de4 on perf-gar-v17-only) also added
the prepareDocumentMatch fast path and the non-KNN adjustDocumentMatch guard;
both are already on master via MB-72489 (#2381) as canFastPrepare() and
adjustKNNDocumentMatch, so this cherry-pick carries only the Reset() change.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
(cherry picked from commit 1b30de4)

With Score="none" and a bounded Size, the request means "return any Size+From
matching documents" — nothing about the result depends on the matches beyond that.
So the collector can stop pulling from the searcher once it has them, instead of
draining the entire match set to count it.

On a 500k-doc corpus this is the difference between scanning every match and
scanning ten: BenchmarkEarlyStopTermTier1ScoreNone goes 18.15ms -> 8.4us and
EarlyStopDisjMin2ScoreNone 41.79ms -> 22.6us on the full branch.

Because the scan really does stop counting, Total becomes a lower bound, and saying
otherwise would silently mislead callers. This adds SearchResult.TotalRelation
("eq" | "gte") to report that, plus the merge rule that any constituent reporting
"gte" makes a merged Total "gte" too.

Gated on the result being independent of unseen documents — no facets (every match
must be counted into the buckets), no KNN (separate hit set), no SearchAfter (the
cursor depends on the full ordering), no nested rollup, no reverse execution, and
sort-by-score only (which degrades to arrival order under score="none"; a field
sort could have its top-k anywhere in the match set).

Two deviations from perf-gar-mod-fix-2's version of this work, both because that
branch's WAND machinery is not present here:

  - TotalRelation is introduced here rather than in 0f56552, which bundled the
    public API with the WANDPruned collector plumbing and the disjunction-searcher
    changes that set it. On this branch the bounded scan is the only producer of a
    lower-bound Total, so the API arrives with its first user and nothing about it
    is dead.
  - index_impl derives TotalRelation from coll.EarlyStopped() alone, not
    "coll.WANDPruned() || coll.EarlyStopped()".

The upstream commit shipped no tests. Added earlystop_test.go, because the
preconditions above are the whole correctness argument and getting one wrong is a
silent wrong-answer bug rather than a slowdown: it checks that a bounded scan still
returns Size hits and reports "gte", that every returned hit is a real match, and
that the scan does NOT engage for facets, field sort, SearchAfter or scoring —
each of which must still report an exact Total.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>