[mtiller]

Steps 0, 1 and 1.5 of the wasm plan: every case exporting as a WebAssembly FMU, in CI, before anything consumes it. Step 2 (the web-worker client that replaces the api/ round-trip) is deliberately not here.

The astro/mystmd refactor is untouched and unblocked — that work is purely additive under web-astro/ and web/, and this is purely additive inside text/.

The native build is a separate pipeline, and that matters

make results runs dvc repro dvc.yaml, which reproduces every stage in that file. The first cut put the 196 build-wasm-* stages there, which quietly made the ordinary book build try to export wasm FMUs with whatever omc it had — and the book-builder image's omc is a release build:

Error: Invalid type of flag simCodeTarget, expected one of None, C, Cpp, ...
       but got the string "wasm-jit".
Error: Unknown option +fmuDirectory.

specs.py now emits them to text/dvc-wasm/dvc.yaml, its own pipeline with its own lock, and make wasm reproduces that file. Neither pipeline can reach the other. text/dvc.lock is consequently byte-identical to refactor.

Everything downstream of make results — plots, Sphinx, ebooks, exes.tar.gz, api/ — behaves exactly as before.

Both export forms, because they are not interchangeable

--simCodeTarget=wasm-jit emits two quite different artifacts and --fmuDirectory is the only difference. One omc run emits both at ~0.7 s per case, so which to ship stays open rather than being decided by the build.

form	count	laid out	median	runs where
wasm/dylink/<case>.fmu	98	8.3 MB	24 KB	model kernel only, binaries/wasm32-om-dylink/. Needs a host that links it against a prebuilt adapter — today an OpenModelica importer. No browser host exists.
wasm/component/<case>.fmu	98	70.5 MB	712 KB	portable fmi-ls-wasm component, binaries/wasm32-wasip2/. What a browser can run today.
wasm/aot/<case>/	98	256.9 MB	2.6 MB	the component transpiled ahead of time into core modules + glue + manifest.json.
wasm/pack/	127 blobs	16.3 MB	—	the AOT content, content-addressed. The deployable layout.

Ahead-of-time transpile (Step 1.5)

A browser cannot instantiate a WebAssembly Component — WebAssembly.instantiate takes core modules. OpenModelica's fmi-simulator page ships jco's transpiler (~9.3 MB) and converts in the page: 1424 ms of a 1598 ms first load (MIC-178). The book's model set is known at build time, so tools/wasm/transpile.mjs does it once per case.

	in-page	AOT
session.load()	1598 ms	169 ms
page open → first result	1854 ms	364 ms
shared download, once	~11.5 MB	2.2 MB

Content-addressed packing

jco names core modules positionally and the numbering shifts per model, hiding how much the cases share. By content they barely differ: 186 MB of core modules is 12.8 MB of distinct bytes, and 1.85 MB of that — the OpenModelica runtime — is byte-identical in every case. fmu.js is 747 KB with four distinct variants.

Per case none of that can be shared, since each aot/<case>/fmu.core9.wasm is its own URL. tools/wasm/pack.py rewrites it:

pack/blobs/<sha256[:16]>.wasm    each distinct core module, once
pack/blobs/<sha256[:16]>.js      each distinct glue file, once
pack/cases/<case>.json           logical name -> blob path, + imports/exports

98 cases → 127 blobs, 16.3 MB — 15.7x, in 0.68 s, hardlinked so it costs no extra disk. The logical names survive because that is what the glue asks for: fmu-core.js builds cores as a Map keyed by jco's name and calls compile(name).

For a reader: ~0.75 MB gzipped for the first figure on a page, then ~16 KB per figure after — against 0.75 MB every time laid out per case. That also sharpens the open Step 0 question: per figure AOT is ~33 KB against dylink's 24 KB, and unlike dylink it has a browser that can run it.

Content-addressing the FMU forms was measured and rejected: it would save 8 KB across all of dylink/ (96 identical 85-byte artifact.json sidecars) and zero across component/.

Two defects fixed, and where the case settings went

gen_wasm.mos emitted setCommandLineOptions("+override=..."), which is not an omc option — it answers Error: Unknown option +override. and carries on — so 8 cases were silently exported with stock parameter values. There is no compile-time equivalent; -override is a runtime flag on the native path too.

Separately, buildModelFMU takes no simulation settings, so an FMU carries the model's own experiment annotation as its DefaultExperiment. For 74 of 97 cases that is not the case's stopTime.

Both belong beside the artifact. specs.py now emits results/cases.json, and make-index.py folds it into results/wasm/index.json — per case, the artifacts plus the stopTime, tolerance, intervals and overrides an importer must apply with fmi3Set* before initialization.

CI

generate_wasm runs in ghcr.io/mtiller/omc-wasm:3998b1f8, pinned to the OpenModelica commit it was built from. No published image has this compiler: --simCodeTarget=wasm-jit needs OM_OMC_ENABLE_RUST=ON, which defaults OFF, and Docker Hub's openmodelica/openmodelica carries release tags only, all pre-wasm. tools/wasm/Dockerfile adds the MSL and toolchain on top of that base and rebuilds in ~35 s; the base is the ~72 minute from-source build on MIC-178.

Two things it needs that only running it revealed:

• clang + lld. IIEV uses external "C" (interpolateVectorTable) with no wasm implementation, so omc compiles the Include sources itself with clang --target=wasm32-wasip1. The wasi sysroot already ships in the base.
• An MSL link. A container job runs with HOME=/github/home, where omc finds no libraries at all, so the workflow links it to the image's MSL before make env.

Gating also changed: generate_results and generate_site were behind the same Full Build label as deploy_site, so a label meaning "build this" also meant "publish to production". Now all three build jobs run on every PR and deploy_site is restricted to push, which the trigger already narrows to master — a PR cannot deploy, with or without a label. labeled comes off the trigger too, since nothing keys off labels now; the Full Build label is unused and can be deleted.

Verification

Last run, all four jobs as intended:

job	result	time
generate_wasm	✅ success	3m
generate_results	✅ success	12m
generate_site	✅ success	1m
deploy_site	⏭️ skipped	—

98/98 cases in all three forms, CI totals matching local byte for byte. Cold 14m30s, warm 3m. Artifact 75 MB at 7-day retention (pack/ is uploaded, aot/ excluded — same content, duplicates removed).