Adopt the WIT mutation schema; retire the byte protocol

[lannbot]

The mutation channel's wire format was a hand-rolled byte encoding whose only specification was prose in a WIT doc comment — an opcode table, a framing grammar and a set of operand encodings, implemented twice by hand (Rust encoder, TypeScript decoder) and kept in agreement by golden vectors and review. This replaces it with an explicit WIT schema and deletes the byte format.

polymorph:dioxus@0.5.0, one export:

export run: async func() -> stream<operation>;

Why now

This branch started as a spike to answer whether the typed schema was affordable. It was not: the typed channel cost ~5 µs per operation, ~5x the byte channel on op-heavy renders and 12x on clear.

That turned out to be almost entirely one runtime's interpreted canonical-ABI lift, not the component model and not the schema. Reported as polyengine#261; upstream landed four fixes; this branch pins 22b5d3d.

op	ops in window	before	after	ratio then → now
create-10k	90 006	4.9 µs/op	0.78 µs/op	5.80x → ~1.8x
create-1k	9 006	5.5 µs/op	0.70 µs/op	4.99x → ~1.6x
append-1k	9 002	5.2 µs/op	0.77 µs/op	1.59x → ~1.1x
clear	10 002	4.1 µs/op	0.15 µs/op	12.02x → ~1.4x

At 1.1–1.8x, on a channel whose cost is a small fraction of the DOM work either way, the maintainability win is worth paying for. The full A/B — methodology, instrumented op counts, cost attribution, and the noise-floor caveats — is preserved in bench/README.md as a historical section.

What this deletes

Deletion is the point of the change; net −3189 lines.

• The wire-format prose in wit/world.wit — opcode table, framing grammar, primitive operand encodings. interface mutations is the schema now, and bindgen owns both sides.
• src/protocol.rs's Batch — opcode constants, utf16_len, the dynstr escape, path/strref encoders, take_frame.
• host/src/decoder.ts — the cursor, decodeBatch's 18-arm switch, FrameDecoder's partial-frame staging.
• tests/vectors.rs and vectors/* — golden vectors existed to keep two hand-written implementations of one format in agreement. There is one generated implementation of each side now.
• The surface-probe fixture's inlined 150-line encoder.

What the schema buys, and what it costs

Beyond the deletion: option<str-ref> replaces the 0xffff sentinel, which frees the reserved id (the interner's guard is now the plain u16 ceiling), and every operand is type-checked at the boundary instead of by a decoder's cursor arithmetic.

The cost is recorded rather than glossed. WIT forbids recursive types — record template-element { children: list<template-node> } is rejected outright ("type template-node depends on itself") — so register-template carries an index arena instead of the natural tree. The arena admits out-of-range and cyclic index graphs a recursive grammar could not express, and host/src/operations.ts validates them. For that one operation, the schema relocates the "two implementations must agree" problem rather than retiring it.

Also: no readDirect (it is stream<u8>-only), so the host pays a copy per batch, and use mutations.{operation} adds an import to the world.

A citation that was wrong in both halves

The guest scheduler's persistent park between renders was justified by the host's parked readDirect session, which the typed channel does not have. The park is legal by a different clause of the same rule: polyengine#162's retention rule is disjunctive — "a retained end, a parked host operation, or an unfinished producer pump" — and mountApp holds the lifted readable end for the instance's lifetime. ("A15" is also no longer a real label; upstream pruned amendment labels, so this cites the contract section and HostActivity instead.)

Testing

Deleting the byte channel deletes the reference implementation of the differential test that proved the typed channel correct — a differential test with nothing to differ against proves nothing. host/tests/operations_test.ts is now an absolute assertion: literal innerHTML at each of eight steps of the counter interaction sequence, captured while the differential test was still green and both channels were proven byte-for-byte identical. The provenance is stated in the test header.

The applyOperations unit tests are unaffected and are now a larger share of the net: arena rehydration including multi-root, rejection of out-of-range and cyclic indices, the non-text attr-value cases, option<str-ref> present and absent, bubbles both ways, create-placeholder.

Gates: cargo check / clippy -D warnings / test, deno task check, deno task test (90), all five examples build and wasm-tools validate clean, and the Playwright lane (8 tests, real Chromium) — which had not been run in this branch before this commit.