[rust-compiler] Store unsupported AST nodes typed instead of via serde_json

[Boshen]

Problem

cargo llvm-lines shows the Rust compiler's generated code is dominated by serde's serde::private::de::content::* machinery — the buffering (de)serializer for the #[serde(tag = "type")] + #[serde(flatten)] Babel-AST enums. In the oxc use case (no Babel/JSON boundary), that serde is reached entirely through one internal round-trip: UnsupportedNode.original_node, stored as serde_json::Value, is serialized in lowering (to_value) and re-parsed in codegen (from_value) to re-emit nodes the compiler does not model.

Change

Replace original_node: Option<serde_json::Value> with a typed enum:

pub enum OriginalNode {
    Expression(Box<Expression>),
    Statement(Box<Statement>),
    Pattern(Box<PatternLike>),
}

• Lowering stores the typed node directly (no to_value).
• Codegen matches the variant (no from_value). PatternLike::as_expression reproduces the exact from_value::<Expression> coercion for the 7 variants PatternLike shares with Expression.
• Adds a hir → ast dependency. The field was opaque JSON specifically to keep HIR decoupled from the AST crate; the coupling is already semantic (the field holds a Babel node), and a generic InstructionValue<N> would be far more churn.

Impact (cargo llvm-lines, dev profile)

crate	total (before → after)	serde (before → after)
react_compiler_reactive_scopes	470,255 → 181,070	248,888 (53%) → 22 (0.0%)
react_compiler_lowering	182,110 → 164,142	38,936 (21%) → 38,900 (24%)

The entire deserialize-side bloat is removed. Lowering's serialize side is unchanged — that is the identifier_loc_index whole-body serialization, a separate follow-up (it carries behavior risk, so it is not bundled here).

Behavior preservation

No behavior change. The old code round-tripped typed → Value → typed, an identity for well-formed nodes, and every original_node is built from a typed node — so the variant already encodes what the type-tag dispatch resolved to. Covered by the rewritten codegen unit tests, the new PatternLike::as_expression tests, and the unknown_statement_lowering integration test.

Verification

• cargo build --workspace, cargo clippy, and all Rust tests for the touched crates pass.
• The corpus-wide e2e fixture comparison (yarn snap --rust) was not run locally: the vendored harness assumes the upstream compiler/ layout and react_compiler_* package names, and the snap runner hits a tsWatch TDZ under Node 24. Please let CI run the full gate.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 966ee9c9d3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Boshen]

Binary size (shipped release index.node — fat LTO + 1 CGU + strip)

Measured the actual stripped release artifact on main vs this branch:

build	bytes	size
main	6,090,416	5.81 MiB
this PR	5,677,632	5.41 MiB
Δ	−412,784	−6.78%

Baseline reproduced byte-for-byte across a clean rebuild, so the delta is solid.

This is the deserialize-side reduction alone. The serialize side (identifier_loc_index's whole-body to_value, ~38.9K IR lines) is still in the binary, so the follow-up that removes it would cut further.