Seeded mix over assoc-shaped data is not reproducible across processes (rand itself is)

[domg123]

Version: 81.1.1 and 81.1.2 (native CLI, macOS arm64; also observed on the WASM build of 81.1.1)

After set_rand_seed, the seeded stream itself is cross-process stable — (rand) sequences repeat exactly. But mix over identically-constructed assoc literals produces different output in different processes, for flat assocs of ~4+ keys and nested assocs alike:

• sorted-key canonicalization of the inputs does not help;
• recursive_matching .false does not help;
• list-shaped inputs do not show the effect;
• within one process results ARE stable (repeated runs in the same process match).

Inferred mechanism: mix's internal traversal of assoc pairs follows a pointer/ASLR-sensitive order, so the seeded stream is consumed in a different order per process. The same effect presumably underlies mix_entities reproducibility reports, but this shows it at the plain code-level mix opcode — anyone moving from string/list-shaped genomes (which replay fine) to assoc-shaped data inherits it silently.

Repro — run this file several times as separate processes and compare (the rand line is byte-identical every run; both mix lines differ run to run; in our testing 6 runs produced 6 distinct flat-assoc outputs):

(seq
  (set_rand_seed "repro-rand")
  (print "rand (stable):      " (rand) " " (rand) "\n")

  (set_rand_seed "repro-mix-flat")
  (print "mix flat assoc:     "
    (format (mix
      {a 1 b 3 c 5 d 7 e 9 f 11 g 13 h 15}
      {a 2 b 4 c 6 d 8 e 10 f 12 g 14 h 16}
      0.5 0.5 {nominal_numbers .false similar_mix_chance 0.5})
      "code" "json" .null {sort_keys .true}) "\n")

  (set_rand_seed "repro-mix-nested")
  (print "mix nested assoc:   "
    (format (mix
      {parts [{part "body" pr 0.9 color {r 10 g 20 b 30}}] traits {speed 0.2 kind "wanderer"}}
      {parts [{part "body" pr 1.1 color {r 240 g 220 b 200}}] traits {speed 0.6 kind "hunter"}}
      0.5 0.5 {nominal_numbers .false similar_mix_chance 0.5})
      "code" "json" .null {sort_keys .true}) "\n")
)

Expected: same seed + same literal inputs → same mix output across processes.
Observed: different output almost every process.