CARD V03 (RF-V) written 2026-09-26 19:40:59 -0700 (clock-read), BEFORE the run. Immutable.
Kind: DIAGNOSTIC TRACE (stock-equal; backend step-pass node snapshots, every iteration; no source change).
Source: src/bipeds_head.c; function ordinal k = 55 of 84 (objfns order). Tool: tools/rfr/stepwalk.py --all (RF-R copy).
Question: after MARK, how many copies of the tail N110..N112 exist before allocation/step #1, which registers each copy
  receives, and at step #2 (cross-jump / tail-dup) which copy hosts the spherical arm's tail.
Predictions:
 P3a: step #1 already shows three tail copies (backend return split, RF-R R3): spherical (after JMP N54's site),
      non-spherical, and the physics fall-through original, each with its own value/pointer registers.
 P3b: ours: the spherical copy carries EDX/ECX in its own allocation (not a clone of the physics copy), i.e. the
      decision is block-local allocation inside the spherical arm, where the EAX/ECX webs of the global-pointer copy
      are still live at the pointer load in list order (pointer load placed BEFORE the copy's last store).
Basis: RF-R R1-R3; V02.

OUTCOME V03 (appended 2026-09-26 19:52:02 -0700; logs/V03_pill_stepwalk.txt; first attempt wrote to a relative /Fo path and CL never ran
  (0 MARK hits); rerun with an absolute work dir):
  Step #1 (post-allocation) already holds THREE inline copies of the tail: spherical N55 value s32@ECX, N56 new pointer
  load s33@EDX, N57 store, then JMP N58 -> L N112; non-spherical N79 value s47@EAX, N80 new pointer s48@ECX, N81 store,
  JMP N82; physics (fall-through original) N109 value s64@ECX, N110 pointer s65@EDX, N111 store, LABEL N112.
  P3a CONFIRMED (the return split makes per-predecessor copies of the MARK tail block before/at allocation).
  P3b CONFIRMED in substance: the spherical copy is coloured on its own (block-local temps, not a clone of another
  copy) and receives ECX/EDX. The deciding fact is therefore WHICH copies exist before allocation: with one
  shared return tail every arm gets its own pre-allocation copy.
