feat(worker): add WorkerDocxodus.prepare() comparison-path warmup

[JSv4]

Problem

createWorkerDocxodus() warms the .NET WASM runtime but does not exercise the comparison engine, so the first compareDocuments() pays a one-time warmup cost (comparison-assembly initialization + JIT of the diff/XML stack) — roughly 2× the steady-state latency. There was no way to pay this ahead of a user action, so downstream apps (crowdsourced-redlines-js, react-docxodus-viewer, …) each reinvented a workaround: ship a pair of seed .docx fixtures and run a throwaway compare purely for the side effect.

Tracking issue in the consumer: JSv4/crowdsourced-redlines-js#2

Fix

Add an optional, idempotent prepare(): Promise<void> on WorkerDocxodus (option 2/3 from the issue — a .NET-side warmup with no caller IO):

• New Warmup() [JSExport] on DocumentComparer runs a real WmlComparer.Compare (+ GetRevisions) against two tiny seed documents built in-memory on the .NET side, forcing the full compare path to resolve and JIT. No seed fixtures shipped, no caller inputs.
• New "prepare" worker message routes to it.
• WorkerDocxodus.prepare() caches a single in-flight warmup promise so repeated/concurrent calls share it, and a compareDocuments() issued while a prepare() is pending does not double-load.

prepare() is never called automatically — skip it and the first compare absorbs the warmup exactly as before.

Note on framing: the issue described the cost as deferred Docxodus.*.wasm/System.*.wasm fetches. In the current build all 40 assemblies are fetched at runtime init, so the ~3s is really first-execution warmup (assembly init + JIT), not network. prepare() amortizes it the same way; the win shows up as timing, and the "no further .wasm fetches after prepare()" guarantee still holds.

Proof (npm/tests/worker-prepare.spec.ts, all green)

Verified via page-level .wasm request monitoring + in-worker performance.now() timing:

Test	Result
After prepare(), a real compare triggers no additional .wasm fetches	✅ loaded: (none)
prepare() makes the first real compare meaningfully faster	✅ cold 1504ms → warm 758ms (~1.9×)
prepare() is idempotent — second call resolves <50ms	✅ first 753ms, second 0.1ms
compareDocuments() while prepare() is in flight does not double-load	✅ 40 requests, 40 unique

Existing worker.spec.ts (create/version/compare) still pass — the harness change is additive.

Surface touched

.NET WASM bridge (DocumentComparer.Warmup) · worker message (docxodus.worker.ts) · proxy (worker-proxy.ts) · types (types.ts) · test harness + new spec · README "First-call warmup" section · CHANGELOG.

Out of scope (per issue)

<link rel=preload> manifest, service-worker caching, and a separate prepareViewer() for the renderer assemblies — follow-ups.

🤖 Generated with Claude Code