gc: adaptive tenuring picks survival age 1 and pays 2-5 unproductive full collections per claude-code turn for it

[proggeramlug]

The adaptive tenuring loop selects survival age S = 1 for a large minority of
minors on claude-code, and each turn pays 2-5 full collections that reclaim
0.2-0.3 % for it. Pinning the age removes them, monotonically. This is on
main; it was surfaced by the regex-header-nursery work (#9845) but is not
caused by it — the unmodified allocator arm shows it identically.

Measurement

One 400-character claude-code reply, offline mock-API rig, sandbox rx.
PERRY_GC_DIAG=1 + PERRY_GC_TRACE=1. A = unmodified allocator,
R = #9845's nursery header. =N pins PERRY_GC_TENURING_SURVIVALS.

arm	tenuring ages chosen	fulls	first-full reclaim	old-gen live at last full
A adaptive	{1:16, 2:6, 3:5, 4:15}	7	0.2 %	49.9 MB
A adaptive	{1:12, 2:6, 3:5, 4:15}	5	0.3 %	48.3 MB
A pinned =2	{2:38}	3	0.3 %	45.3 MB
R adaptive	{1:13, 2:5, 3:5, 4:11}	5	0.2 %	48.7 MB
R adaptive	{1:15, 2:5, 3:4, 4:12}	7	0.2 %	46.6 MB
R pinned =2	{2:35}	3	0.3 %	45.5 MB
R pinned =4	{4:36}	2	0.2 %	34.7 MB

Two readings:

• The effect is on main, not on perf(regex): allocate the RegExp header in the nursery, not the malloc arm #9845. A and R adaptive overlap completely
  ({7,5} against {5,7}) and their age distributions are near-identical, so the
  allocator change does not alter tenuring behaviour. A pre-registered
  prediction that R would show more fulls was falsified by this matrix.
• Fewer fulls, monotone in the pinned age, in both arms: 5-7 adaptive -> 3
  at =2 -> 2 at =4, with old-gen live at the last full falling 48-50 MB ->
  45 MB -> 35 MB.

Why this is a defect and not a tuning preference

The fulls being removed reclaim 0.2-0.3 %. By #9589's rule — price a
collection by what it freed, never by occupancy — they are unproductive, and the
policy is scheduling 2-5 of them per turn because it tenured objects that were
about to die. Promoting on the first survival converts a short-lived object into
old-gen occupancy, and occupancy is what schedules the full.

This is the cost half of the argument; the arena lane owns the producer and the
structural fix (the occupancy rule's S = 1 + desired/influx reaching exactly 1
with no rung at 2 or 3, and the invariant that S = 1 destroys the measurement
that could refute it, since prev_copied = 0 holds the guard false forever).
Its rule — occupancy may lower only to 2, and reaching 1 belongs to the
mortality-measuring exits — is consistent with this matrix.

Three constraints on how these numbers may be used

1. Count claim only. This box ran at load 100-300 throughout; its CPU
   figures are unusable and none are quoted here. Every number above is a count
   or a byte total from the trace.
2. Not merged with the other host. These fulls reclaim 0.2-0.3 %; the same
   arms on a quiet host reclaim 14-25 %. The two hosts sit in different regions
   of the same policy and their numbers are not combined into one table.
3. PERRY_GC_TENURING_SURVIVALS=2 is the diagnostic, not the fix. It is
   documented diagnostic-only, and moving a threshold is explicitly not the
   remedy being proposed.

Reproduce

stream_scale.py LABEL HOME PORT 100 400 150 \
   --env PERRY_GC_DIAG=1 --env PERRY_GC_TRACE=1 --stderr OUT.txt -- <binary>
stream_scale.py LABEL HOME PORT 100 400 150 \
   --env PERRY_GC_DIAG=1 --env PERRY_GC_TRACE=1 \
   --env PERRY_GC_TENURING_SURVIVALS=2 --stderr OUT_t2.txt -- <binary>

Count collection_kind == "full" records, and read tenuring_survivals= off
the [gc-copy-minor] ran lines.