codegen: reconcile the root-spill threshold (#8623, 32M) with the RS4GC instruction budget (#8586, 1.57M) — budget-gap functions get refused instead of spilled

[proggeramlug]

Problem

Two #8583 mechanisms disagree on where a function is "too big for RS4GC":

• perf(codegen): raise root-spill default to the measured RS4GC fan-out cliff (#8620) #8623 set the root-spill decision threshold to 32,000,000 relocations (DEFAULT_ROOT_SPILL_RELOCATIONS), measured at the point RS4GC's rewrite still finishes in bounded time.
• fix(codegen): delete the unsound pre-RS4GC optnone knob; add a post-rewrite instruction budget and per-unit codegen stats (#8583) #8586 refuses a unit when a function's post-rewrite instruction count exceeds 1,572,864 (PERRY_LL_RS4GC_MAX_INSTRS), the point the -Os/-O3 optimizer goes super-linear.

Since a function's added instructions ≈ its relocation estimate, a function whose estimate lands in (1.57M, 32M) does not spill (estimate < 32M) yet overruns the budget (post-rewrite > 1.57M) → the budget assertion hard-refuses the whole module.

Concrete case

Compiling the Claude Code cli.js 2.1.112 bundle, __AnonShape_ec12be8bd7b0a2b9_constructor (a large closed-shape object-literal constructor) grew 34,009 → 2,280,128 instructions under RS4GC — above the 1.57M budget, below the 32M spill threshold — and was refused. It only compiled with PERRY_ROOT_SPILL_RELOCATIONS=1500000. (Estimate accuracy for this shape is fixed by #8678, which counts property/index stores; but even with an accurate ~2.25M estimate it still won't spill at the 32M default.)

Options

1. Align the spill default with the budget — set DEFAULT_ROOT_SPILL_RELOCATIONS at/below PERRY_LL_RS4GC_MAX_INSTRS so anything that would overrun the budget spills first. (Note this revisits perf(codegen): raise root-spill default to the measured RS4GC fan-out cliff (#8620) #8623/codegen: #8589 root-spill default threshold (4M) too low — spills moderate-fan-out functions and slows their compile #8620: the "8M fans-out-fine" cases codegen: #8589 root-spill default threshold (4M) too low — spills moderate-fan-out functions and slows their compile #8620 measured would also overrun the 1.57M budget, so they should spill — RS4GC finishing ≠ the optimizer handling the result.)
2. Preferred: make the budget trigger a spill-retry, not a hard refusal — when a function overruns the budget post-rewrite, re-lower it with request_shadow_frame_spill() (roots to the shadow frame, RS4GC skips it) instead of refusing the module. This is the "fix: make compile output TypeScript-developer friendly #8421/fix(codegen): unbreak in-process RS4GC on inline asm and relocation-grown functions #8128 root-lowering choice, never refuse" principle applied to the budget, and it needs no threshold tuning at all.

Option 2 is the durable fix. Found while getting the cli.js bundle to a working native binary (#8583).