GC: 50 MB of an idle TUI's old gen is swept-dead but never returned — the idle reducer's cycle cannot compact by construction

[proggeramlug]

Half the old gen of an idle claude-code TUI is dead, swept, and unreachable to the OS, because nothing compacts it and the reducer's cycle cannot.

Measured

Compiled cc TUI (cli_2.1.112.js, macOS), a session with 279 collections behind it, idle after a paste, PERRY_GC_DIAG=1:

[gc-idle-reclaim] start attempt=1 external_collections=279 old_in_use=93608664
[gc] blocks: general=42 (1 live), non_general=143 (130 live, survivors+longlived+old), freed_bytes=2321504
[gc-old-free] reusable_bytes=49986912
[gc-idle-reclaim] done old_in_use=93608664->93606360 reclaimed_old=2304 freed=2370304

50.0 MB of the 93.6 MB old gen is on the free list — already swept, already dead, reusable for future old-gen allocation but never returned. 130 live blocks before the cycle, 130 after. A second reducer full a few seconds later moved the free list to 50.1 MB and released nothing.

This is the residue #9589's Finding 2 was pointing at (it attributed it to pacing defaults; the census in #9637 corrected the mechanism for the allocator half, and this is the arena half).

Why the idle reducer cannot fix it

Old-page defrag is the only mechanism that empties partially-occupied old blocks, and it rides on a moving minor:

• gc/policy.rs: evacuation_policy_allowed = !low_pause_non_moving, and low_pause_non_moving = progress_kind.is_budgeted().
• The idle reducer's cycle is budgeted by construction (it is stepped in 4 ms wake-checked slices), so evacuation is off in it by definition.
• GcCycleState::new_full takes no old-page selection at all; only new_minor_fallback does. A full mark-sweep never compacts, budgeted or not.

So no amount of gate-tuning in gc/idle_reclaim.rs reaches this memory. What would: a moving minor with select_old_page_defrag_pages armed, started from the park hook when the free-list residue crosses a threshold.

What that needs decided first

1. gc: PERRY_GC_OLD_DEFRAG=0 (the rollback path for default-on old-page relocation) is exercised by no test and no CI arm #7917. Defrag went opt-in (PERRY_GC_OLD_DEFRAG=1) pending "the reproducer and a dependency-scale stress corpus are clean", and that corpus still does not exist — none of the 19 benchmark programs can even produce a candidate page (selection wants dead_bytes >= live_bytes). An idle cc TUI does produce them, at 50 MB, which makes it the missing corpus as much as it makes it the beneficiary.
2. Pause. Idle is not a promise: the user can end it with a keystroke at any moment. A moving minor is not sliceable the way the budgeted stepper is, so this needs a measured pause budget and a policy for what happens when a wake lands mid-evacuation.

Filed off #9589's reducer work (PR #9643 fixes the pricing bug that made the reducer switch itself off before it ever got near this).

https://claude.ai/code/session_011qE5TRRJFzqxN44K34AnG2

[proggeramlug]

Fixed by #9671, which landed on main via a merge train. The train was merged as its own branch, so the close-keyword in #9671 never fired — closing manually. Verified the fix is on main by patch-id (git cherry).