The default rack is now Pattern-only (31c7c01), but membership filters SELECTION,
not SPAWNING - so all 13 unselected guns still ran `predict` + `spawnBullets`
every tick to feed fitness tables nobody reads. Measured waste: Tsetlin alone
0.98 ms/tick, KNN 0.30, plus 10 more. This generalises the gate TMPATTERN already
had to every gun, behind `TR_VBULLET_ADMIT_ONLY` (default 1 = gate, 0 = old).
OFFLINE COST (release build, 400 ticks, min of 2 reps, 13-gun rack):
OFF 2.14 ms/tick (implied 467 ticks/s)
ON 0.04 ms/tick (implied 27667 ticks/s)
-> reclaimed 2.10 ms/tick, ~98% of the virtual-bullet cost. Tsetlin's 0.98
disappears, KNN's 0.30 disappears, only Pattern (0.017) survives.
Note the absolute scale is lower than an earlier unoptimised measurement (~6.1
ms/t) because this is a -d:release build; the ON-vs-OFF DELTA is the robust result.
LIVE TICK RATE (ONE frozen binary, 4 runs/arm x 4 rounds, all 8 concurrent, gate
varied by env only):
ON 146.2 ticks/s (142.3, 145.2, 147.9, 149.4)
OFF 86.9 ticks/s ( 95.0, 41.4, 99.1, 112.1)
NO OVERLAP: ON min 142.3 > OFF max 112.1. Excluding a game-outcome outlier in
the OFF arm, OFF max is still 112.1. Outside the noise.
**+68% tick rate**, which also makes every future A/B faster. Both arms still pay
the fixed 8192-slot ring scan in tickBullets.
SAFETY, verified not assumed: Pattern is admitted under the shipped rack, so its
own fitness keeps accumulating and the SELECTED gun is unchanged - Pattern 100% in
all 4 runs both arms, and Pattern was the ONLY gun with vShots>0 in the ON arm
while all 13 had vShots>0 in the OFF arm. So admission is the correct predicate.
Mid-round transitions are safe by construction: only predict/spawn are gated, while
tickBullets still resolves every active bullet and the feedback case still calls
the owning gun's onResult.
A TOOLING BUG THIS CAUGHT, and a correction to the task's assumption:
`acceptance_offline_vs_online.nim` IS affected (I had assumed it was not). It
compares the live per-gun vShots against an offline replay that always spawns all
guns, so under the default gate the online non-Pattern vShots are 0 while offline
is ~400 - a guaranteed mismatch. Fixed by pinning `TR_VBULLET_ADMIT_ONLY=0` inside
that test (same putEnv/defer pattern as TR_RECORD_WORLDSTATE), keeping 12/12. The
test is about offline/online METRIC parity, so it needs every gun spawning.
Unaffected (verified from source): audit_virtual_guns.nim,
measure_cornering_guns.nim, sweep_tm_pattern.nim - all offline, none read
GUN_STATS_PATH.
Guards: test_vbullet_admit_gate 12 (new, pure), test_gun_harness 39,
test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41,
test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 28,
test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20,
test_tm_pattern_learning 3, acceptance_offline_vs_online 12/12. ModularBot compiles.
SirStone
3142b70aa5