Building a list by pushing into an ArenaVec reallocates within the arena as the vec grows. Child AST nodes are allocated between pushes, so the growing buffer is almost never the arena's most recent allocation — every doubling allocates a new buffer and abandons the old one as permanent dead space in the arena. A list of N elements costs ~⌈log2 N⌉ grow events and ~2-4× its final size in arena bytes.

This PR builds every list in a scratch std::Vec instead, and moves the completed list into the arena with a single exact-sized allocation (a memcpy).

Design

• One scratch vec per list element type (ScratchBuffers, ~27 fields). Nested lists with the same element type (statements of nested blocks, arguments of nested calls) share one vec, used like a stack via watermarks: record the vec's length at list start, push elements, drain your own tail at completion. List parsing is strictly LIFO — parse errors are stored rather than thrown, and element closures cannot early-exit their caller — so a list always drains its own elements before its enclosing list resumes. (This is the scratch pattern used by Zig's std.zig parser, split per element type since our lists are heterogeneous.)
• The buffers are checked out of a thread-local cache when the parser is created and returned emptied when it is dropped (panics included, via a drop guard). They are never freed, staying grown to high-water capacity across parses on the same thread — once a thread is warm, list building performs no heap allocation at all. The only unsafe is an empty-struct lifetime cast at checkout/checkin.
• The cursor.rs list helpers carry the discipline for almost all sites; comma-separated lists with no closing token get new parse_separated_list/parse_separated_list_from helpers. A debug assert at parse end catches any future site that pushes without draining (checked by every conformance run, which uses debug-assertions = true).

Results

just allocs (Arena allocs, Sys allocs, Sys reallocs are all unchanged):

The residue is mostly lexer StringBuilder escape paths and the comments vec, which are not converted.

Parser benchmarks improve up to ~10% (parser/App.tsx, parser/react.development.js, estree/checker.ts) and regress nowhere; conformance output is byte-identical across all suites.

🤖 Generated with Claude Code