Document post-#55 jemalloc rejection

[aksOps]

What changed

• Adds the missing public decision record for Stream A's post-Stop holding the whole tree's references: cold RSS 830,612 kB to 286,780 kB #55 jemalloc measurement.
• Keeps the system allocator; the effective product-code and dependency diff is empty.
• Preserves the older measurement that favored jemalloc and records why Stream B must remeasure the pinned Node tree before its final RSS gate.

Why

The N-033 owner ruling requires the rejected measurement to remain public instead of disappearing with Stream A's product revert. The two measurements describe different binary memory models: the older result predates #55's retained-reference redesign, while Stream A measured the post-#55 build.

Measured result

Both release binaries scanned Kubernetes commit 7fbcad5e8ec387e23061602eb2d4b110b049410c three times from clean stores under taskset -c 0,1 and /usr/bin/time -v.

allocator	median wall	median peak RSS
system	132.20 s	285,324 kB
jemalloc 0.7.0	115.65 s	320,808 kB
delta	-16.55 s (-12.52%)	+35,484 kB (+12.44%)

All six store SHA256 values were identical. Stream A therefore rejects jemalloc because RSS regressed even though wall time improved. Stream B still owns the exact pinned-Node allocator comparison required by the ruling.

Movement and ratchets

• Graph rows moved: 0
• Edges gained: 0
• Reclassifications: 0
• Baselines changed: 0; no arthron gate --rebase was needed
• Pins changed: 0; no arthron pin --write was needed

Verification

• Effective branch diff: docs/decisions.md only
• git diff --check: pass
• cargo fmt --all --check: pass
• Prior adversarial batch: 29/29 gates, 15/15 pins, and 15/15 whole-row joins; it found the measured RSS rejection
• Prior final batch after the allocator revert: fmt, clippy, full corpus suite, 29/29 gates, and 15/15 pins passed
• Current origin/main: CI, Corpus gates, and CodeQL passed
• All branch commits and the merge commit are SSH-signed, single-line, and trailer-free

The owner reopened A only to retain the public measurement. A third heavy batch was not run because the per-stream protocol permits exactly two; the current effective change is documentation-only.