Make per-call nested function instantiation allocation-free

[lahma]

Summary

Every call of a function whose body declares inner functions re-instantiates those functions (spec-required — each call's closures capture that call's environment). ETW profiling shows the instantiation chain (FunctionDeclarationInstantiation → InstantiateFunctionObject → ScriptFunction..ctor) at 4.6% of program time on closure-producing workloads. This makes each instantiation cheaper — instantiating a named sloppy nested function drops from ~7 side allocations (~250–300 B) to zero (only the ScriptFunction object itself remains, +16 B for two new lazy fields).

Five commits:

1. Lazy name/length/prototype descriptors — a shared pending-descriptor sentinel materializes the real PropertyDescriptor on first read (sharing actual descriptors across instances is provably unsafe: ValidateAndApplyPropertyDescriptor mutates them in place). The sentinel's value accessors throw so a leak fails loudly. The per-definition-constant function name JsString is cached on JintFunctionDefinition.
2. Lazy sloppy-mode arguments/caller poison descriptors — a plain sloppy function no longer creates any property dictionary at all; delete-then-redefine demotes to dictionary preserving key-resurrection order; initial key order unchanged; %ThrowTypeError% resolution stays deferred exactly as before.
3. FDI hoisting-loop trims — FunctionsToInitialize becomes a readonly record-struct array (was LinkedList) with the binding Key precomputed at BuildState; the intrinsic lookup is hoisted out of the loop; SetFunctionName's per-call JsString allocation is skipped for named declarations.
4. Pre-size the FDI var-binding dictionary (mirrors the adjacent lexical-declarations pattern) instead of growing per insert.
5. Materialize the pending prototype in [[Construct]] only when consulted.

Functional verification beyond suites: 77 checks (descriptor shapes before/after calls via getOwnPropertyDescriptor, key order, delete/redefine/freeze, loop-closures, NFE recursion, generators/async, Proxy, bind, with, class) — output byte-identical to a pristine main build.

Benchmarks (default jobs, adjacent A/B on latest main)

Benchmark	main	PR	Δ
FunctionDeclarationAlloc DeclareMany	61.98 µs / 168.79 KB	52.57 µs / 129.73 KB	−15.2% time, −23.1% alloc
FunctionDeclarationAlloc ConstructEach	202.07 µs / 418.79 KB	199.68 µs / 395.35 KB	−1.2% time, −5.6% alloc
ClosureCallBenchmarks (5 rows)	—	—	neutral (SloppyEmptyClosureCall settled by launchCount-5: 25.49 ±0.82 → 25.00 ±0.39 ms, allocations byte-identical)
SunSpider recursion rows (top-level functions — lane not exercised)	—	—	neutral

Gates

• Jint.Tests: 3,569 (net10.0) + 3,506 (net472) passed, 0 failed
• Jint.Tests.PublicInterface: green both TFMs (descriptor surface exercised)
• Jint.Tests.CommonScripts: green both TFMs
• Test262: 99,429 passed, 0 failed (run twice)

🤖 Generated with Claude Code