Final retest after double-payment fix (commit bffe8a48), 2026-08-19/20.

gpt-4o-mini, index mode, REPS=5:
  small: solvedRate=100% avgTokens=1271 (was 1770 pre-fix, 0% two fixes ago)
  large: solvedRate=100% avgTokens=3351 (was 5210 pre-fix, 0% two fixes ago)

qwen3:14b, index mode, REPS=5:
  small: solvedRate=100% avgTokens=2903 (was 40% pre-fix)
  large: solvedRate=60% avgTokens=5204 (was 60% pre-fix, unchanged)

index_capped (cap=8, no discover-tools) diagnostic, gpt-4o-mini, REPS=5:
  small: solvedRate=0% -- target tool (registered last) never survives the
    naive registration-order truncation; cap has no relevance ranking.
  large: solvedRate=0% -- same cause, confirmed via debug trace showing
    the target tool in the "8 more" overflow line every time.

full mode sweep, gpt-4o-mini, n=3/size, diverse-domain fillers:
  n=15: solvedRate=100% avgTokens=1495
  n=30: solvedRate=100% avgTokens=2243
  n=45: solvedRate=100% avgTokens=3047
  n=60: solvedRate=100% avgTokens=3621
  (never degrades up to 60 tools; generic "filler-N" names instead of
  diverse domain names produced a false 0% at n=15 in an earlier flawed
  attempt at this sweep -- semantically null names give the model nothing
  to rule out, so it thrashes trying each one; this is a probe-design
  confound, not a finding about full mode itself, corrected before citing)
