[drietsch]

Shrinks the paged.web engine wasm (bin/blitz_web_bg.wasm) by 1.7 MB with a dedicated build profile, and tightens its size gate so growth is caught.

What changes

• packages/web-render/Cargo.toml: a [profile.wasm-release] profile (inherits = "release", lto = "fat", codegen-units = 1). The default release profile has no cross-crate LTO and 16 codegen units. Native tests and benches keep their profiles.
• scripts/build-wasm.sh and scripts/build-wasm-perf.sh build with --profile wasm-release. wasm-opt -Oz stays.
• Size gate: build-wasm.sh now reads its ceiling from the manifest's capabilities.wasm[blitz].maxBytes, so the two can't drift. That value drops from 64 MiB to 9 MiB (9 437 184 bytes), about 4.7% above the new size.
• @paged-media/web goes to 0.1.0-canary.18. The engine sources changed (Cargo.toml) and so did the manifest, so package-hash.mjs --check needs the bump. Merging publishes canary.18.
• Docs: architecture.md and the ADR 308 row in docs/README.md now give the new ceiling. status.md and the perf record get the measured sizes.

Size

Built the way the publish workflow builds it: build-wasm.sh --engine with Rust 1.93.0, wasm-bindgen 0.2.125 (from Cargo.lock) and binaryen 119.

	raw bytes	brotli
before (default release profile)	10 723 939	2 612 273
after (wasm-release)	9 016 823	2 496 172
delta	−1 707 116 (−15.9%)	−116 101

Built in CI for this PR (the vitest job): 9 015 481 bytes (brotli 2 497 009). The published 0.1.0-canary.17 wasm is 10 723 633 bytes, so CI to CI the difference is −1 708 152 bytes. Local and CI builds differ only by a few hundred bytes of embedded build paths.

No slowdown

I timed the five perf workloads from tests/common/workloads.rs in node against both wasm builds through the exported entries. Each run did 10 alternating rounds of 40 iterations per workload on one machine, and the table gives the median of the round medians:

workload	before	after	delta
article-40, 1 frame	5.16 ms	4.84 ms	−6.2%
article-40, flow into 4	8.90 ms	8.82 ms	−0.8%
article-40, flow into 12	21.40 ms	21.12 ms	−1.3%
table-300, 1 frame	12.18 ms	12.16 ms	−0.2%
styled-runs-200, 1 frame	2.10 ms	2.03 ms	−3.2%
cold boot (instantiate)	9.9 ms	8.7 ms

For every workload, the JSON output of the two builds is byte-identical. test/perf/perf-budgets.spec.ts with PERF_SHOW=1 (3 interleaved runs each) gives the same door, wasm-call and byte counts for both builds, and its wall-clock stays within noise.

Tests

• REQUIRE_REAL_ENGINE=1 pnpm -r exec vitest run against the new wasm: web-bundle 339/339, web-conformance 103/103, web-model 290 passed (the 9 corpus-dependent tests skip, as in CI). Contract-import lint and typecheck are clean.
• cargo fmt --check, clippy -D warnings for all three feature sets, cargo test --features blitz, cargo test and the perf_budgets count budgets all pass.
• The gate fails on the old 10.7 MB wasm and passes on the new one.

🤖 Generated with Claude Code