perf(codegen): async-to-generator boxes are never cleared — completed activations retain 96% of asyncpipe's young generation #7933

Summary

96% of asyncpipe's young generation is retained by activations that have already completed.

The async-to-generator transform boxes every body local of every async function, and nothing ever frees or clears a box. Every local of every completed activation therefore remains a permanent GC root for the life of the process. On asyncpipe that is ~48,000 completed activations pinning their entire local sets.

Measurement

A temporary one-env-var probe made box roots rewrite but not mark, i.e. kept them valid without treating them as live:

The real live young set is ~6,600 objects. At that size a minor costs ~0.8 ms instead of 62 ms.

Supporting attribution: crate::box::scan_box_roots_mut is 92% of all pointer roots and 89.5% of root-scan time on this program. Essentially every box slot is a from-space pointer, so visiting it during CopyingMark evacuates the object it names — 93,380 of minor #1's 156,236 evacuations come straight out of box roots.

Suggested direction

Clear an activation's boxes at its terminal state — do not free them.

Clearing keeps BOX_REGISTRY monotonic and every box address permanently valid, so both perry#4898's rejection semantics and #7906's positive pointer cache are untouched. Freeing would break both.

Why this is filed rather than fixed

It is a codegen change in the async-to-generator transform, which CLAUDE.md lists as a known-weak area ("it boxes every body local into a shared mutable cell typed Any"), and the terminal-state hook has to be right for every exit path — normal return, throw, and early termination — or a live local gets cleared while still reachable. That is a silent-use-after-clear, not a crash.

Relationship to other issues

• gc: asyncpipe collects at 1200-1650 ns/object, including a 122 ms minor that handled zero objects #7915 is closed by this: its two stated premises were counter artefacts (see that issue for the correction). This is the real mechanism behind the same symptom.
• gc: incremental old-gen work costs 14% on a program that never collects (asyncpipe, zero GC cycles) #7909's 12% depends on collection being affordable on this program; a ~95% cut in young-set size is the affordability.
• bench: asyncpipe@120 sits exactly on a nursery trigger boundary and is not a usable A/B target #7926: the trigger cliff exists because the retained set puts the program right at a boundary.

Full reproduction and per-scanner attribution: gc-handoff/ROOTS-NOTES.md.