===== boot A (--kv bypass)
  [0] code           REF      16384tok  wall=  6.99s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [0] code           REF2     16384tok  wall=  6.84s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [0] code           STD      16384tok  wall=  7.07s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [0] code           INT8     16384tok  wall=  6.66s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [0] code           INT8G1   16384tok  wall=  7.10s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [1] code           REF      16384tok  wall=  6.85s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [1] code           REF2     16384tok  wall=  6.85s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [1] code           STD      16384tok  wall=  7.11s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [1] code           INT8     16384tok  wall=  6.68s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [1] code           INT8G1   16384tok  wall=  7.10s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [2] wiki           REF      16384tok  wall=  6.85s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [2] wiki           REF2     16384tok  wall=  6.86s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [2] wiki           STD      16384tok  wall=  7.12s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [2] wiki           INT8     16384tok  wall=  6.68s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [2] wiki           INT8G1   16384tok  wall=  7.10s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [3] wiki           REF      16384tok  wall=  6.87s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [3] wiki           REF2     16384tok  wall=  6.88s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [3] wiki           STD      16384tok  wall=  7.15s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [3] wiki           INT8     16384tok  wall=  6.69s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [3] wiki           INT8G1   16384tok  wall=  7.22s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [4] math           REF      16384tok  wall=  7.03s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [4] math           REF2     16384tok  wall=  7.00s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [4] math           STD      16384tok  wall=  7.35s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [4] math           INT8     16384tok  wall=  6.84s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [4] math           INT8G1   16384tok  wall=  7.31s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [5] math           REF      16384tok  wall=  7.05s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [5] math           REF2     16384tok  wall=  7.10s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [5] math           STD      16384tok  wall=  7.29s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [5] math           INT8     16384tok  wall=  6.87s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [5] math           INT8G1   16384tok  wall=  7.26s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [6] longctx        REF      16384tok  wall=  7.28s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [6] longctx        REF2     16384tok  wall=  7.28s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [6] longctx        STD      16384tok  wall=  7.44s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [6] longctx        INT8     16384tok  wall=  6.99s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [6] longctx        INT8G1   16384tok  wall=  7.41s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [7] longctx        REF      16384tok  wall=  7.28s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [7] longctx        REF2     16384tok  wall=  7.19s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [7] longctx        STD      16384tok  wall=  7.64s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [7] longctx        INT8     16384tok  wall=  7.00s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [7] longctx        INT8G1   16384tok  wall=  7.48s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [8] sci            REF      16384tok  wall=  7.19s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [8] sci            REF2     16384tok  wall=  7.18s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [8] sci            STD      16384tok  wall=  7.35s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [8] sci            INT8     16384tok  wall=  6.90s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [8] sci            INT8G1   16384tok  wall=  7.22s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [9] sci            REF      16384tok  wall=  7.07s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [9] sci            REF2     16384tok  wall=  7.10s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [9] sci            STD      16384tok  wall=  7.29s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [9] sci            INT8     16384tok  wall=  6.86s  finish=length  leg=0F/3200S acc=3200x16/0x32
  [9] sci            INT8G1   16384tok  wall=  7.27s  finish=length  leg=0F/3200S acc=3200x16/0x32
[i8] kernel census: int8 calls=64000 rows=131072000 reconstructs SKIPPED=64000 fallback (non-4bpw) calls=0
[i8] excluded projections: none  excluded calls=0
[i8] per-arm census (int8 kernel calls): {'REF': 0, 'REF2': 0, 'STD': 0, 'INT8': 32000, 'INT8G1': 32000}
  prefill positions aligned: [8, 8, 8, 8, 8, 8, 8, 8, 8, 8] per prompt, identical across all 5 arms
saved per-arm prefill logits+hidden to /cache/walk/A.pt
EXIT=0
===== boot B (--kv bypass)
[i8] teacher forcing from /cache/walk/A.pt: arm REF,REF2,STD,INT8,INT8G1 of a bypass boot, 10 prompts
  [0] code           REF      16384tok  wall=  7.16s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [1] code           REF      16384tok  wall=  7.01s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [2] wiki           REF      16384tok  wall=  6.98s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [3] wiki           REF      16384tok  wall=  7.03s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [4] math           REF      16384tok  wall=  6.99s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [5] math           REF      16384tok  wall=  6.93s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [6] longctx        REF      16384tok  wall=  6.91s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [7] longctx        REF      16384tok  wall=  6.90s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [8] sci            REF      16384tok  wall=  6.93s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [9] sci            REF      16384tok  wall=  6.88s  finish=length  leg=3200F/0S acc=3200x16/0x32
[i8] kernel census: int8 calls=0 rows=0 reconstructs SKIPPED=0 fallback (non-4bpw) calls=0
[i8] excluded projections: none  excluded calls=0
[i8] per-arm census (int8 kernel calls): {'REF': 0}
  prefill positions aligned: [8, 8, 8, 8, 8, 8, 8, 8, 8, 8] per prompt, identical across all 1 arms
saved per-arm prefill logits+hidden to /cache/walk/B.pt
EXIT=0
===== boot C (--kv k4v4)
[i8] teacher forcing from /cache/walk/A.pt: arm REF,REF2,STD,INT8,INT8G1 of a bypass boot, 10 prompts
  [0] code           REF      16384tok  wall=  7.12s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [1] code           REF      16384tok  wall=  7.13s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [2] wiki           REF      16384tok  wall=  7.11s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [3] wiki           REF      16384tok  wall=  7.13s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [4] math           REF      16384tok  wall=  7.22s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [5] math           REF      16384tok  wall=  7.31s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [6] longctx        REF      16384tok  wall=  7.18s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [7] longctx        REF      16384tok  wall=  7.14s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [8] sci            REF      16384tok  wall=  7.13s  finish=length  leg=3200F/0S acc=3200x16/0x32
  [9] sci            REF      16384tok  wall=  7.29s  finish=length  leg=3200F/0S acc=3200x16/0x32
[i8] kernel census: int8 calls=0 rows=0 reconstructs SKIPPED=0 fallback (non-4bpw) calls=0
[i8] excluded projections: none  excluded calls=0
[i8] per-arm census (int8 kernel calls): {'REF': 0}
  prefill positions aligned: [8, 8, 8, 8, 8, 8, 8, 8, 8, 8] per prompt, identical across all 1 arms
saved per-arm prefill logits+hidden to /cache/walk/C.pt
EXIT=0
