kit tokens per kit-arm run (full size/4 of every distinct kit .md file opened; upper bound)
model                                          T1    T2    T3    T4    T5    T6    T7    T8    max  >20k  first=compose
sonnet                                      9577  9275 13989  5944  6251  5520 10314  2838  13989     0          8/8
opus                                        1274  9853  6802  2838  6251  1274  1274  1274   9853     0          8/8
gpt-6-luna                                 18259 15053 17129  9876 13289  8260 12131 16701  18259     0          8/8
gpt-6-sol                                  25245 21571 26404 13289 13289  8260 12131 18817  26404     3          8/8
gemini-3.8-flash-medium                    26534 20725 10808  9357  9357  8260  8198 18539  26534     2          8/8
opencode_muse-spark-1.3-contributor-free   22943 20001 18437 25958 49929 41516 47132 16272  49929     6          0/8

Note (Muse add-on): OpenCode loads skills through its own `skill` tool, which this parser does not count. Muse called
`skill` first in its runs (checked on T1-kit), so "first=compose 0/8" is a parser gap, not a routing failure, and
Muse's numbers omit its SKILL.md loads. Its reference reads (up to ~50k) are real. Panel rows are unaffected.