# Chaos benchmark scenes (perf-p3; scripts\chaos_bench.ps1, notes/perf-plan.md "Chaos benchmark"): the most chaotic
# moments reachable by warp cards (notes/warp.md) and time-based scripted input - boss fights, brawls, fire, rain,
# debris, the heaviest hub views. Columns (|-separated):
#   name | episode | job | flags (direct, play, make=..., env=K=V) | seconds | pad | window | rec frames
# pad: time-based tokens after START (runtime/window.c r<s>); GPT / HPT / TPT macros (fight / hub walk / turret, *n
# repeats; GPT = 12.25 s, HPT = 20.6 s) are expanded by chaos_bench.ps1; REC marks where -Record starts the GS recording
# (a no-op in benchmarks). START comes ~50-58 s after the start (the title's intro); direct jobs then load and play
# their intro cutscene. window: from-to in seconds of the run's wall clock (PS2_FPS_TRACE) over which fps and stages
# are summarised (scouted with -Scout, 2026-10-24: the action is on screen there). rec frames: VBlanks recorded by
# -Record (default 240). Recordings: build\gsrec\chaos_<name>.gsrec (never committed).
# Not kept: m1_boss (3-minute intro), m4_brawl (Sly climbing, no fight), m7_boss (fight after ~150 s), turret / engine /
# TNT jobs (long Binocucom briefings, then the pause menu opened under the scripted input), m5_battlezone (load stalls
# under parallel runs; the corpus has its gas cloud: gpu-M8 m5_battlezone).
m1_brawl | 1 | m1_brawl | direct | 200 | r30:none,GPT*2,REC,GPT*10 | 100-195
m2_boardroom | 2 | m2_boardroom_brawl | direct | 200 | r30:none,GPT*3,REC,GPT*10 | 65-195
m3_boss | 3 | m3_boss | direct | 200 | r30:none,GPT*2,REC,GPT*10 | 108-195
m5_boss | 5 | t5_heist_boss_1_intro | direct | 200 | r30:none,GPT*4,REC,GPT*8 | 130-195
m8_boss | 8 | m8_boss | direct | 200 | r12:none,GPT*1,REC,GPT*12 | 56-88
ep1_hub | 1 | m1_brawl | play | 200 | r5:none,HPT,REC,HPT*7 | 85-195
ep3_hub | 3 | | play | 200 | r5:none,HPT,REC,HPT*7 | 70-195
ep5_hub | 5 | | play | 200 | r5:none,HPT,REC,HPT*7 | 70-195
