sim: fork PushShapes into frozen per-version packages (v1/v2/v3)

[ElmoPA]

Physics changed twice without the data recording which version produced it, so
historical episodes could not be reproduced. A v2-collected u_socket episode
replays to coverage 0.000 under v3 friction (stored 0.953) -- that divergence is
why 1384 episodes had to be regenerated, and it was only discoverable by hand.

Fork each era into a frozen package so any dataset can be replayed under the sim
that made it:

v1 git HEAD geometry L15=8 L20=24 L25=20 L29=52 no u_socket
v2 pre-v3 snapshot L15=3 L20=3 L25=54 L29=54 solid_pusher
v3 current same geometry + pocket-bottom-only socket friction

Verified by replay (Tsimulation/examples/replay_zarr.py):
u_socket_3000_v2 under v2 -> 0.953 vs stored 0.953 reproduces
u_socket_3000_v3_replacements under v3 -> 0.956 vs stored 0.975 reproduces
u_socket_3000_v2 under v3 -> 0.000 vs stored 0.953 diverges

Also:

• pushshapes/init exposes get_env/get_module/available_versions and aliases
  the historical module paths in sys.modules, so all 12 existing importers keep
  working unchanged and see the same module objects (constants included).
• v3 stamps sim_version into episode_init, at the env layer so every collector
  inherits it. New data is self-describing instead of era-inferred from which
  optional keys happen to be present.
• replay_zarr and mouse_collect take --sim-version; teleop no longer silently
  inherits whatever CURRENT_VERSION happens to be.
• Removed redundant Tsimulator/ duplicate trees and 86 untracked *.bak files.
  Content preserved: Tsimulator/sim_v1 == v1/ and sim_v2 == v2/, byte-identical.

Known remaining: Tsimulation_legacy/ (v1 duplicate) is still live with 16
referrers via the PUSHSHAPES_SIM env-var switch; that mechanism and this registry
are not yet reconciled.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_029adb0d-ec6d-4577-a617-65f29c6bc23a

[ElmoPA]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat(robot): deploy graph policies on Python 3.10 #578
• feat(pipeline): sync current inference graph runtime #577
• feat(pipeline): sequential H-Net pipeline, flat stages instead of a nested trunk #567
• feat(rldb): heterogeneous u_socket action support #566
• refactor(config): organize all Hydra configs by family and task #551
• tsim: sim env + viz/scripts + physics tuning #549
• sim: consolidate to Tsimulation/{sim_v1,sim_v2}, retire v3 numbering #546
• sim: fork PushShapes into frozen per-version packages (v1/v2/v3) #542 👈 (View in Graphite)
• feat(rldb): h264 video-backed images + role-in-key annotations + span selection #541 : 1 other dependent PR (#550 )
• TTT-Linear across-time layer (RoboTTT-style): closed-form inner SGD on fast weights, per-episode reset, tanh-gated ~identity at init; DualTrunkLevel ttt knob; arms fix2ada+ttt (H-Net) and txar_sdpada(+ttt) flat pair #540
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.