feat(measure): component-aware harness with meld + code-section deltas (v0.9.1 PR-R)

[avrabe]

Summary

User-requested fairness pass on the v0.9.0 measurement harness:

1. Component-aware comparisons — wasm-opt can't process components directly; use meld fuse to produce a comparable core form, then run wasm-opt + LOOM on that.
2. Code-section deltas — the right metric for optimizer effectiveness. File bytes include attestation / debug / custom sections that aren't the optimization target.
3. Per-run timeout — concurrent harness invocations were hanging on the 2.3 MB calculator melded core for >60 min each; bounded.

Key finding

The original code_section_bytes was returning 0 for every fixture because it parsed wasm-tools dump output where the code-section size is LEB128-encoded (0a ab 06 = 811 bytes), but the parser grabbed the first integer from the address column (0x3ab → 0). Every prior code-section number in the report was silently zero. New parser uses wasm-tools objdump which prints decimal byte counts in a clean pipe-separated table.

Honest results matrix

File size (total bytes):

Workload	Baseline	LOOM	wasm-opt -O3	LOOM Δ%	wasm-opt Δ%
gale	1,941	1,846	1,925	−4.9%	−0.8%
calculator_root	2,337,724	2,327,794	(errors on components)	−0.4%	n/a
simple_component	261	212	(errors)	−18.8%	n/a
calc_component	442	392	(errors)	−11.3%	n/a

Code section (optimizer-relevant):

Workload	Baseline	LOOM	wasm-opt	LOOM Δ%	wasm-opt Δ%
gale	811	795	795	−2.0%	−2.0%
calculator_root	106,017	96,515	n/a	−9.0%	n/a
simple_component	9	9	n/a	0%	n/a
calc_component	33	33	n/a	0%	n/a

Components via meld (fused-core baseline):

Workload	meld baseline	wasm-opt	LOOM	wasm-opt Δ%	LOOM Δ%
calculator_root	2,294,685	1,927,848	(timeout)	−16.0%	n/a
simple_component	90	90	41	0%	−54.4%
calc_component	135	135	86	0%	−36.3%

What the new data shows

1. On gale's code section LOOM and wasm-opt are TIED at −2.0%. LOOM's 4-point file-size advantage came from removing non-code custom sections, not from better code optimization. Honest narrative now possible.
2. calculator_root's 106 KB code section dilutes to 0.4% at file level (2.3 MB). The code-only view shows LOOM does −9.0% there — significant.
3. Components have tiny code sections (9 B, 33 B); after meld fuses adapter glue, the post-meld core is real code (90 B, 135 B) and LOOM gets −54.4% / −36.3%.
4. wasm-opt beats LOOM at scale on the calculator core (−16% post-meld), but LOOM dominates on small adapter-heavy modules. The crossover is somewhere between 135 B and 2.3 MB.

Code changes

• scripts/measure_corpus.sh: with_timeout wrapper, component magic-byte detection, meld fuse integration, fixed code_section_bytes (LEB128 bug), 15-field row schema with code-section columns, two report tables.
• docs/measurements/v0.9.0-corpus-baseline.md: regenerated with real numbers.

🤖 Generated with Claude Code