perf: reuse selected reconstructed deltas

Revert "Enable rolling repetition penalty by default"

Enable rolling repetition penalty by default

pbc3: use int16 canvases

pbc3: streamline decoder patch application

pbc3: restore PIL resampling and fix benchmarks

pbc3 refactor

pbc3: promote simplified codec

homemade numba rng + resampling

palette merge into ops

Final refactor hopefully, 7 files + 1 model

some more cleanup

removed dead code

some more cleanup + may be faster now

cleanup

last resort benchmark check

refactor revert to check

release version, quick benchmark mode to verify

heuristic backup for when learned model crashes during patch fill

PBC3.0 release version pre-test

full rl pipeline phase 2: rl encoder prepared working teacher oracle for distillation

full rl pipeline phase 1: search candidates refactor

rl training environment added

animation + sweep fixes, preset update (removed high compression, made compression the new balanced)

internal model implemented, tested, gonna test qualitatively

base values double read fix

two new features downsample warmup and search pipeline shortcuts

preload more tbh

preload numba

front page improvement

server fixes compress stream etc

pbc3.0 first update