compat-eslint: lazy parent — defer parent resolution to first .parent read

[johnsoncodehk]

Summary

materialize(tsNode, ctx) previously walked up the TS parent chain and built every ancestor top-down so the leaf could be returned with a fully-resolved parent. Rules that never read .parent (e.g. JSX-attribute counters, type-checker leaf reads) paid for the entire chain anyway.

LazyNode.parent is now a lazy getter backed by a PARENT_UNSET sentinel; first read calls resolveParent, which walks the TS chain incrementally — one ancestor per .parent step. Bottom-up materialize constructs ONLY the requested leaf with parent: undefined. Top-down construction is unchanged — slot getters still pass this as the parent, so the constructor sets _parent directly and .parent reads stay O(1).

Three details that fell out of the design

• Routers read tsNode.parent.kind directly. Routers that previously checked parent._ts.kind (BindingElement, Method/Get/Set accessor) now read the TS parent — the same node either way for these kinds, and routers must work without a resolved parent under the lazy-leaf path.
• convertChild re-points a cached child's _parent when it's still PARENT_UNSET. Each TS node has exactly one ESTree parent, so the slot owner who's calling us IS that parent. Without this, ChainExpression's expression slot leaves the inner Identifier pointing at PARENT_UNSET and resolveParent drills to the wrong wrapper. Caught by prefer-rest-params regressing on arguments?.length.
• Parent getter prefers a side-effect parent over the resolveParent return value. Wrappers like ChainExpression occupy the same TS slot as their inner expression, so resolveParent returns the outer wrapper — but the inner slot's convertChild already wrote the correct intermediate parent via the setter.

Module-level currentCtx lets routers call convertChild with parent: undefined and still resolve ctx for the cache lookup. Set on entry to materialize, restored on exit.

Numbers (Dify cold lint, 3 runs)

	master	lazy	delta
Wall	13.55–13.88s	10.96–11.85s	~18% faster
User CPU	~21s	~17.5s	~17% faster
Node count	174,367 (60 kinds)	138,046 (48 kinds)	~21% fewer

JSXAttribute 24k→16.6k, JSXElement 15k→10.4k, JSXOpeningElement 8k→4.6k — those JSX ancestors are now built only when the rule actually reads .parent.

Net −206 lines in lazy-estree.ts.

Test plan

• lazy-estree.test — incl. ~67k-assert dogfood parity sweep
• predicate-coverage 152/152
• bench 107 rules × 33 files — clean parity
• dogfood 107 rules × 30 files — 0 divergences, 0 crashes
• scope-compat 24/24, selector-analysis, ts-ast-scan, compat-pipeline, jsx-react-x PARITY ✓
• Dify cold lint — 3 runs, ~18% wall improvement, 21% fewer nodes