perf: use 4096 as default SINDI window size

[chasingegg]

Summary

• change the default SINDI window size from 65535 to 4096
• preserve the supported 1024-65535 range and explicit overrides

Related to #1730.

Motivation

Local IP benchmarks on NQ, FEVER, HotpotQA, and MS MARCO SPLADE data measured
window sizes 1024, 4096, 16384, and 65535 at topK 10, 100, and 1000. Window
4096 had the highest QPS in every dataset/topK combination for both tested
dim_max_score_ratio values (1.05 and 0.756), with unchanged available recall.

Window 4096 balances cache locality and per-window overhead: its 16 KiB score
buffer fits comfortably in L1 and enables finer-grained pruning and earlier
TopK threshold updates, while avoiding the extra metadata and cursor/loop work
seen with window 1024.

Query-level counters on the same workload confirm the pruning effect (ranges
across the four datasets):

topK	Scored postings: 4096 / 65535	Documents scanned: 4096 / 65535
10	51.11-70.49% / 72.08-85.25%	1.94-5.07% / 17.67-29.33%
100	68.21-83.69% / 92.40-97.75%	11.07-22.85% / 63.51-82.10%
1000	89.28-96.52% / 99.92-99.99%	49.48-71.63% / 99.28-99.94%

SPLADE IP performance (dim_max_score_ratio=1.05)

The benchmark used the first 1,000 real queries, no filter, three timed rounds
per point (median reported), and eight search threads on an 8-core Neoverse-V2
machine. Table values are QPS; the best window in each row is bold.

Dataset	topK	window 1024	window 4096	window 16384	window 65535	4096 vs 65535
NQ SPLADE	10	1532.49	1835.77	1299.12	1142.87	1.606x
NQ SPLADE	100	1195.56	1390.65	966.98	898.52	1.548x
NQ SPLADE	1000	892.42	1031.54	800.34	817.60	1.262x
FEVER SPLADE	10	588.10	648.83	473.66	419.98	1.545x
FEVER SPLADE	100	495.12	538.15	391.07	354.08	1.520x
FEVER SPLADE	1000	397.57	439.65	340.67	330.22	1.331x
HotpotQA SPLADE	10	596.53	662.29	487.48	434.45	1.524x
HotpotQA SPLADE	100	504.97	553.05	407.37	368.87	1.499x
HotpotQA SPLADE	1000	410.59	456.33	358.61	349.60	1.305x
MS MARCO SPLADE	10	707.98	991.16	613.29	530.09	1.870x
MS MARCO SPLADE	100	549.13	736.46	461.38	404.31	1.822x
MS MARCO SPLADE	1000	411.35	533.73	358.53	357.22	1.494x

All tested window sizes produced the same sampled result hash for each
dataset/topK slice, and all available recall values were unchanged.

Test plan

• built the Release knowhere_tests target from a clean upstream/main worktree
• knowhere_tests "[config]"

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: chasingegg

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details

Needs approval from an approver in each of these files:

• OWNERS [chasingegg]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[mergify]

@chasingegg 🔍 Important: PR Classification Needed!

For efficient project management and a seamless review process, it's essential to classify your PR correctly. Here's how:

1. If you're fixing a bug, label it as kind/bug.
2. For small tweaks (less than 20 lines without altering any functionality), please use kind/improvement.
3. Significant changes that don't modify existing functionalities should be tagged as kind/enhancement.
4. Adjusting APIs or changing functionality? Go with kind/feature.

For any PR outside the kind/improvement category, ensure you link to the associated issue using the format: “issue: #”.

Thanks for your efforts and contribution to the community!.

[chasingegg]

/kind improvement

[alexanderguzhva]

/lgtm