Add CI; fix bench/ops.ts against the current mountApp signature

[lannbot]

bench/ops.ts was calling mountApp with the pre-source option shape:

mountApp({ componentBytes, translator, root, onError })   // stale
mountApp({ source: { componentBytes, translator }, root, onError })  // current

Every op failed all three attempts with Cannot use 'in' operator to search for 'plan' in undefined. Because bench_worker.ts deliberately converts a failed op into medianMs: null rather than aborting the run (one op failing shouldn't kill the whole table), complete breakage rendered as a table of N/A rather than as a crash.

The type checker already knew

deno check bench reports it outright:

TS2353: Object literal may only specify known properties,
        and 'componentBytes' does not exist in type 'MountOptions'.

Nothing was running it. deno.json's check task covered host/src host/tests and stopped there. Widened to include bench. The three translator: unknown parameters are now UntranslatedArtifacts["translator"], taken from the embedder's own type so the call site stays pinned to what mountApp accepts.

CI

The hole surviving is the more general problem: there was no CI at all, so every gate was "someone remembers to run it". This adds .github/workflows/ci.yml, which runs the existing just recipes unchanged — new gates belong in the justfile, not in the workflow.

Two things it deliberately does:

• Builds the fixture and counter components. host/tests/counter_test.ts and host/tests/fullstack_test.ts throw when their .wasm is missing rather than skipping, so these steps are load-bearing: drop one and the suite goes red, not quietly green.
• Runs both halves of the cross-language protocol gate. cargo test's golden_matches_committed pins the Rust encoder to vectors/; deno task test's decoder_test.ts pins the TypeScript decoder to the same committed files. Neither was running automatically before.

Toolchain versions (Rust, Deno, wasm-tools, just) are pinned, matching how the justfile pins POLYENGINE_REV and Cargo.toml pins dioxus/wit-bindgen. A floating toolchain can turn clippy -D warnings red on a new lint with no commit to blame.

Verified locally

Ran the workflow's exact sequence: just deps (17s from a clean checkout), just check, just fixtures, just example counter, just test — cargo 10 passed + golden_matches_committed ok, deno 87 passed including the 4 golden-vector tests. bench/bench_worker.ts remove-row stream now returns a real median instead of null.

Independent of the stream<u8> vs stream<operation> spike (#18), which stays as it is.