The numbers

All numbers: Apple M1 Max, criterion, deterministic corpora — means with 95% CI unless noted. Single-machine: compare relatively. Method and caveats on the previous page.

Current full-suite baseline (2026-08-31)

Bench	Mean	95% CI
value_encode	409.44 ns	[408.99, 411.51]
value_decode	872.73 ns	[872.36, 876.42]
hnsw_build_2k_64d	125.92 ms	[125.73, 126.11]
hnsw_search_2k_64d	19.04 µs	[18.97, 19.10]
hnsw_build_pq_2k_64d	365.69 ms	[363.03, 369.07]
hnsw_search_pq_2k_64d	37.79 µs	[36.85, 39.14]
pq_train_2k_64d (parallel)	64.93 ms	suite context
pq_train_10k_128d (parallel)	381.68 ms	suite context
bm25_exact_2k	8.033 ms	[7.968, 8.143]
bm25_indexed_2k	490.10 µs	[483.54, 498.24]
dot_768d	81.07 ns	[80.42, 81.68]
l2_768d	84.26 ns	[82.48, 84.98]
cosine_768d	224.29 ns	[220.86, 231.60]
edge_link_10k	392.98 ms	suite context
edge_delete_sweep_100	49.70 ms	suite context
delete_heavy/delete_half_2p5k	230.39 ms	suite context
delete_heavy/insert_unique_scalar_5k	185.72 ms	suite context
compound_prefix_scan/eq_leading_only_5k	251.09 µs	suite context
selective_window_verify/eq_50_of_5k	236.20 µs	suite context
selective_window_verify/eq_500_of_5k	619.56 µs	suite context
order_by_indexed_5k/asc_limit20	312.61 µs	suite context
order_by_indexed_5k/desc_limit20	1.009 ms	suite context
create_text_index_ondisk_5k	230.22 ms	suite context
create_vector_index_ondisk_2k_8d	402.82 ms	suite context

Named deltas vs the prior baseline: bm25_exact_2k +5.0% (the CJK-aware tokenize pass on a latin corpus — documented residual, inside the guard); the delete/edge family sits above its recorded isolated AFTERs (the suite-vs-isolated context gap, unchanged code, isolated re-probes confirmed); value_decode +3.0% and the distance kernels +1–5% are ambient machine drift. No regression attributable to code; the guard holds.

The optimization program’s headline deltas

Each row’s BEFORE is the recorded pre-change baseline (provenance in the engine’s BENCHES.md; conventions on the overview).

Verify-candidates batching (dense indexed windows verify with one ordered walk instead of per-key point-gets):

Bench	BEFORE	AFTER	Δ
eq_500_of_5k (10% density)	745.14 µs	576.25 µs	−22.7%
eq_50_of_5k (1% density)	221.21 µs	229.82 µs	+3.9% (point-gets kept; within guard)

Sort indexes (filterless order_by over a complete scalar index):

Bench	BEFORE	AFTER	Δ
asc_limit20	2.642 ms	289.5 µs	−89.0%
desc_limit20	2.795 ms	1.024 ms	−63.4%

Compound prefix-only windows (the all_docs_indexed flag):

Bench	BEFORE	AFTER	Δ
eq_leading_only_5k	1.540 ms	232.41 µs	−84.9% (6.6×)

Edge adjacency (O(degree) delete cascades via endpoint-first derived namespaces):

Bench	BEFORE	AFTER	Δ
edge_delete_sweep_100	241.0 ms	40.5 ms	~5.9× — O(degree) vs O(E)
delete_half_2p5k	566.8 ms	194.5 ms	~2.9×
edge_link_10k	273.5 ms	359.4 ms	+40% — RATIFIED

The ratified trade: two extra rows per link buy O(degree) cascades — deletes were the workload-blocking hazard. The alternative shapes (consolidated per-endpoint values, pure-lazy adjacency) were measured and rejected with numbers.

Parallel PQ training (identical codebooks, bit-for-bit):

Bench	Sequential	Parallel	Δ
pq_train_2k_64d	177.2 ms	67.4 ms	2.6×
pq_train_10k_128d	1465.0 ms	356.2 ms	4.1×

Graph reads, endpoint-direct (parity verdict — kept for the shared layout, not speed):

Bench	BEFORE	AFTER	Δ
hub_out_knows (313 rows)	37.08 µs	37.12 µs	+0.1%
hub_in_knows	29.58 µs	29.96 µs	+1.3%
traverse_hub_2hops	570.5 µs	581.4 µs	+1.9%

Next: scaling.