feat(stylex): add native StyleX compiler and pass

[MarcelOlsen]

Adds native StyleX support to OJ's development and build pipelines, using the clean-room Rust compiler fru vendored at crates/stylex. The current compiler source is synced from lovable@d61e617fb93d174d892bbffb0dc83e2290c39508; the OJ workspace manifest remains adapted for this repository. Babel 0.19.0 is the compatibility oracle.

Development transforms OJ's parsed AST before lowering. Compiled rules travel with cached modules, populate the stylesheet registry, replace @stylex; in served CSS, and trigger CSS updates when gated modules change. The build pipeline uses a Rolldown string transform with a splice source map, then assembles and inserts the stylesheet after bundling. The pass auto-enables when the app depends on @stylexjs/stylex; the stylex config section, --stylex-config, or OJ_STYLEX_CONFIG supplies explicit settings.

{ "stylex": { "include": ["src/**"], "useCssLayers": true, "dev": true } }

The latest compiler sync reduces temporary allocations by iterating simple evaluated namespaces directly and representing ordinary single-property paths implicitly. Complex styles retain the existing validation and flattening pipeline. The library's RuleRegistry also shares prepared rules when repetition warrants it, returns to inline storage when repetition disappears, and avoids duplicate preparation on first emission. OJ currently calls the one-shot assembler, so those registry savings require a later integration change. The standalone CLI queue change is outside this vendored library.

Compiler measurements on a 3,103-module corpus, Apple M5 Max, Rust 1.95.0, one worker:

Production compilation metric	Before	After
Time inside create()	33.56 ms	28.86 ms
Allocation/reallocation calls	2,511,299	2,341,595
Cumulative allocated bytes	482.86 MB	464.89 MB

Total compiler instructions fell about 3%. The allocated-byte reduction is temporary allocation traffic; peak heap stayed unchanged. Stage timing and allocation counting used separate runs. Whole-OJ experiments using a separate AST-hook prototype found 0.7–1.5% fewer instructions with PGO, but no consistent wall-time improvement. Those are prototype measurements, not a claimed speedup for this PR's string build integration. The AST-hook adapter and whole-OJ PGO setup are separate work; native scope analysis remains opt-in through FRU_NATIVE_SCOPES=1.

The retained compiler snapshot produced byte-identical complete outputs before and after the optimizations across four corpora at one and eight workers and passed 5,325 live Babel comparisons in its source repository. That repository keeps the full conformance harness and integration fixtures. Tests in this sync pin numeric property order, the $$css insertion position, dynamic class paths, error ordering, and the registry's transition from repeated to unique rules.

Two known integration constraints remain:

• OJ's experimental disk module cache is off by default. Its StyleX configuration salt and rule replay do not validate filesystem resolution dependencies before a hit is accepted. A local restart probe reproduced stale output after changing only the package name. Correct invalidation must account for configuration, aliases, and resolved module identity; imported theme contents should not become importer dependencies. The compiler has dependency validation, but this host cache does not yet use it.
• Babel 0.19.0's malformed @position-try output is reproduced for compatibility. StyleX-substituted sheets therefore use Lightning CSS error recovery; other sheets remain strict.