Track candid performance series and benchmark deltas

[sasa-tomic]

Summary

This issue records the performance work on sat-perf-improvements and the benchmark evidence for the full series relative to ba72cf4.

The series currently includes eight focused changes:

Original series (5 commits):

• fast-path small Nat / Int encode-decode paths
• avoid cloning record field queues during decode
• fast-path exact primitive vector decode
• remove duplicate whole-file type rebuilding in candid_parser typing
• skip trivia-only tokenizer bookkeeping when trivia capture is disabled

New additions (3 commits):

• bulk memcpy for primitive vector encoding on LE platforms (14.7x encoding speedup)
• batch cost tracking for primitive vector decode (41% decode speedup)
• Int::decode small-value fast path (12% decode speedup + 50% heap reduction)

Benchmark Comparison (Combined)

Baseline: ba72cf4 (master)
Compared branch state: sat-perf-improvements (all 8 commits)
Tool: canbench (wasm32)

Key improvements

Benchmark	Metric	Baseline	Optimized	Change
vec_nat	Encoding	1.10B	66.54M	-93.9% (16.5x)
vec_nat	Decoding	1.49B	917.08M	-38.5%
vec_nat	Total	2.59B	983.62M	-62.0%
vec_nat32	Encoding	136.28M	16.79M	-87.7% (8.1x)
vec_nat32	Decoding	1.06B	406.87M	-61.6%
vec_nat32	Total	1.20B	423.67M	-64.7%
vec_nat64	Encoding	161.45M	33.57M	-79.2% (4.8x)
vec_nat64	Decoding	1.06B	411.07M	-61.2%
vec_nat64	Total	1.22B	444.64M	-63.6%
vec_int16	Encoding	123.7M	8.4M	-93.2% (14.7x)
vec_int16	Decoding	1.07B	411.07M	-61.5%
vec_int16	Total	1.19B	419.48M	-64.8%
option_list	Decoding	28.35M	23.70M	-16.4%
option_list	Heap	2 pages	1 page	-50%
variant_list	Decoding	27.05M	22.07M	-18.4%
variant_list	Heap	2 pages	1 page	-50%
btreemap	Encoding	4.70B	526.00M	-88.8% (8.9x)
btreemap	Decoding	15.74B	13.41B	-14.8%
btreemap	Total	20.44B	13.94B	-31.8%

Full benchmark table

status	name	master	optimized	ins Δ%	HI master	HI opt	HI Δ%
-	blob	6.33M	6.33M	+0.0%	66	66	+0.0%
-	btreemap	20.44B	13.94B	-31.8%	1179	1154	-2.1%
-	btreemap::1. Encoding	4.70B	526.00M	-88.8%	159	159	+0.0%
-	btreemap::2. Decoding	15.74B	13.41B	-14.8%	1020	995	-2.5%
-	extra_args	3.51M	2.90M	-17.4%	0	0	+0.0%
-	nns	26.64M	25.56M	-4.1%	2	2	+0.0%
-	nns::0. Parsing	17.86M	16.84M	-5.7%	2	2	+0.0%
+	nns::1. Encoding	2.09M	2.09M	+0.0%	0	0	+0.0%
-	nns::2. Decoding	5.74M	5.69M	-0.9%	0	0	+0.0%
-	nns_list_proposal	77.38M	74.12M	-4.2%	17	17	+0.0%
+	nns_list_proposal::1. Encoding	6.88M	6.89M	+0.1%	3	3	+0.0%
-	nns_list_proposal::2. Decoding	70.50M	67.23M	-4.6%	14	14	+0.0%
-	option_list	36.38M	24.35M	-33.1%	2	1	-50.0%
-	option_list::1. Encoding	8.03M	641.58K	-92.0%	0	0	+0.0%
-	option_list::2. Decoding	28.35M	23.70M	-16.4%	2	1	-50.0%
+	text	12.09M	12.09M	+0.0%	99	99	+0.0%
-	variant_list	35.28M	22.71M	-35.6%	2	1	-50.0%
-	variant_list::1. Encoding	8.22M	636.51K	-92.3%	0	0	+0.0%
-	variant_list::2. Decoding	27.05M	22.07M	-18.4%	2	1	-50.0%
-	vec_int16	1.19B	419.48M	-64.8%	261	195	-25.3%
-	vec_int16::1. Encoding	123.70M	8.40M	-93.2%	261	130	-50.2%
-	vec_int16::2. Decoding	1.07B	411.07M	-61.5%	0	65	—
-	vec_nat	2.59B	983.62M	-62.0%	172	151	-12.2%
-	vec_nat::1. Encoding	1.10B	66.54M	-93.9%	33	33	+0.0%
-	vec_nat::2. Decoding	1.49B	917.08M	-38.5%	139	118	-15.1%
-	vec_nat32	1.20B	423.67M	-64.7%	518	387	-25.3%
-	vec_nat32::1. Encoding	136.28M	16.79M	-87.7%	518	258	-50.2%
-	vec_nat32::2. Decoding	1.06B	406.87M	-61.6%	0	129	—
-	vec_nat64	1.22B	444.64M	-63.6%	1031	771	-25.2%
-	vec_nat64::1. Encoding	161.45M	33.57M	-79.2%	1031	514	-50.1%
-	vec_nat64::2. Decoding	1.06B	411.07M	-61.2%	0	257	—

Notes

• All optimizations preserve the wire format — byte-identical serialization output.
• Forward and backward compatibility verified via round-trip tests for all affected types.
• Zero regressions across all 12 benchmarks (minor noise on nns_list_proposal::1. Encoding).
• The btreemap improvements come from the Nat fast paths (earlier commits); the Int fast path further helps option_list and variant_list.
• Heap reduction in option_list/variant_list comes from avoiding BigInt allocation for small Int values.
• vec_nat encoding speedup (16.5x) comes from the Nat::encode u64 fast path avoiding BigUint arithmetic for small values.
• vec_nat32/vec_nat64 encoding speedups come from bulk memcpy on little-endian platforms.
• vec_nat decoding (38.5% improvement) benefits from Nat::decode u64 fast path; further gains possible with reduced per-element serde overhead.
• Heap increase entries marked "—" indicate the baseline had 0 pages (all heap was used for encoding), so the optimized branch's decode-only heap is not comparable.

Benchmarks added in #718

New benchmarks (vec_nat, vec_nat32, vec_nat64) added to master separately to track these types going forward.