Four of the five adversaries imported the OLD package (tankroyale_botapi
1.0.1); only SittingDuck used robocode_tankroyale_botapi 1.0.7, which is what
the rest of the repo requires. A previous report claimed OscillatorBot was
already on 1.0.7 - that was WRONG, and OscillatorBot turned out to crash the
MOST (8 SIGSEGVs in the first reproduction, 15 in its historical /tmp logs).
THE CRASH, reproduced with an identical stack in every case:
botThreadEntry -> run -> adversary run -> go -> dispatchPendingEvents ->
tankroyale_botapi-1.0.1/event_queue.nim(89) addEvent -> realloc/rawDealloc ->
SIGSEGV
Counts, old API: 60 melee battles x 8 rounds gave RandomMover 1, PatternMover 3,
WaveSurfer 0, OscillatorBot 8; 6 battles x 6 rounds vs SittingDuck gave 4/2/0/3.
ROOT CAUSE: the main->bot event hand-off. 1.0.1 passes a lock-protected
seq[BotEvent] (signalTick writes gPendingEvents, dispatchPendingEvents copies it
under lock). 1.0.7 uses a Channel[seq[BotEvent]] (send(move(pending)) /
tryRecv). The old path copied string-bearing BotEvent payloads across threads
every tick, churning ORC refcounts on the shared heap until the freelist was
corrupted. 1.0.7's own source documents this as the gdb-confirmed fix.
WHY IT MATTERED MORE THAN IT LOOKED: the crash silently corrupted measurements.
Against a stationary duck, crash contamination inflated WaveSurfer's rest
fraction from 12.4% (clean) to 20.7%; in a focused run the server logged
'Bot left: OscillatorBot' while the game continued and its score stopped
growing. So every gauntlet run tonight was fighting adversaries that were
partially dead - which is a second, independent reason the user's instinct that
these bots were bugged was correct, and why they should not be used as a
measurement baseline. (The per-gun REAL hit rates are unaffected: those came
from DrussGT battles.)
FIX: all four migrated to robocode_tankroyale_botapi 1.0.7. NO API adaptations
were needed beyond the module rename - every symbol these bots use is identical
in 1.0.7, verified by diffing the two packages (constants/utils/json_parse/
schemas semantically identical; the movement and intent procs in bot.nim are
byte-identical). The .nimble files now require robocode_tankroyale_botapi.
VERIFIED: 120 melee battles x 8 rounds plus 6x6 vs SittingDuck -> 0 SIGSEGV in
all four stderr logs (0 bytes). Behaviour unchanged: sub-1% absolute drift in
mean speed, rest fraction, reversal rate, mean range and perpendicular fraction,
all within run-to-run spread; the one >=3-sigma flag (WaveSurfer perpendicular
relative to DrussGT) was isolated against a stationary opponent and shown to be
the chaotic closed loop, not the migration. test_wavesurfer_velocity passes 7/7.
NOT migrated, reported only: GotoTest_garage, OscillatorBot_garage (archived
copy), PPO_Bot_garage, QBot_garage, SAC_LSTM_Bot_garage - older experiment
garages, left alone deliberately.
SirStone
d3b3c28cdf