fix(v8): accept UTF-8 labels and decode UTF-16LE

[thisisanubhav]

Description of Changes

Fixes #5994.

The V8 host previously rejected new TextDecoder('utf8') and new TextDecoder('utf-16le'). Both are constructed eagerly by the h3-js@4.5.0 browser bundle, so the dependency could not finish module evaluation even though its UTF-16 decoder is unused.

• Recognize the WHATWG labels for UTF-8 and UTF-16LE, trim ASCII whitespace, normalize case, and report the canonical encoding name.
• Add non-streaming UTF-16LE decoding in the existing JavaScript builtin: explicit little-endian byte reads, surrogate pairs, replacement/fatal errors, leading BOM removal, and buffer-view offsets. A pending high surrogate plus an odd trailing byte produces one EOF error.
• Accept omitted decode input as empty bytes.
• Add a V8 host regression module reproducing dependency-style eager construction and covering aliases, decoding, malformed sequences, options, shared buffers, odd view offsets, repeated calls, and large input.

Compatibility: this remains a subset of TextDecoder. UTF-16BE and other encodings, streaming, and ignoreBOM: true remain unsupported. The existing native UTF-8 decode path (including its existing BOM behavior) is unchanged. UTF-16LE is implemented in JavaScript, without a new native API or dependency. The separate CLI success-exit behavior after module-evaluation errors is outside this fix.

API and ABI breaking changes

None. Previously rejected labels now work; the public host ABI is unchanged.

Rollback safety impact

n/a

Expected complexity level and risk

2. Limited to the V8 text-encoding builtin and its regression test. UTF-16LE performance has not been benchmarked. Native host execution still needs upstream CI validation as described below.

Testing

• Passed Prettier 3 checks for both JavaScript files, RUSTUP_TOOLCHAIN=stable rustfmt --check --edition 2024 crates/core/src/host/v8/mod.rs, and git diff --check.
• Passed the checked-in regression module in Node v25.8.1 V8 using a temporary vm.SourceTextModule harness (node --experimental-vm-modules /tmp/issue-5994-check.mjs). Only the unchanged native UTF-8 primitives were shimmed; UTF-16LE executed the production implementation.
• Passed 84,025 UTF-16LE differential cases against Node's TextDecoder in both replacement and fatal modes: every zero-, one-, and two-byte input, surrogate/BOM boundary combinations, and deterministic random inputs.
• Evaluated the unmodified h3-js@4.5.0 browser bundle in that context: latLngToCell(37.775938728915946, -122.41795063018799, 9) returned 8928308280fffff; gridDisk(cell, 1) returned seven cells. Master reproduced the utf8 failure and, after changing only that label, the utf-16le failure.
• Native host test not run locally: the offline Cargo attempt stopped before compilation because the workspace's iai-callgrind Git dependency was not cached. Disk was critically low (256–416 MiB free; ENOSPC constraint), so no further dependencies were downloaded or large builds attempted. Formatting used the installed stable toolchain (Rust 1.97.1), rather than the pinned 1.96.1 toolchain.
• Upstream CI/reviewer check: cargo test --locked -p spacetimedb-core text_decoder_dependency_construction_and_decoding. No standalone publish/build flow was run locally.

[CLAassistant]

All committers have signed the CLA.

[coolreader18]

I'm going to close this in favor of #6045, though I might nab the test file if that's okay.

[coolreader18]

@thisisanubhav would you please sign the CLA?

[thisisanubhav]

@coolreader18 CLA signed