Add SO2 embedding modes and init-conditioned embedding interface for RL

[ht0324]

Summary

This PR adds SO2 embedding experiment infrastructure for RL and fixes embedding inference to use initial-angle conditioning (matching estimation training behavior).

Changes

• Added RL observation modes:
• raw
• embed_only
• concat (raw + embedding)
• Added SO2 RL configs for hidden sizes 64 and 128:
• embed_only_so2_h64.yaml
• embed_only_so2_h128.yaml
• concat_so2_h64.yaml
• concat_so2_h128.yaml
• Updated embedded config to explicitly set observation_mode: embed_only.
• Extended embedding interface to accept optional init_pos during inference.
• Updated Reacher wrapper to extract reset-time SO2 init encoding and pass it during embedding computation.
• Added SO2 hidden-size sweep configs/scripts:
• rnn_so2_h64.yaml
• rnn_so2_h128.yaml
• scripts/train_so2_rnn_hidden_sweep.sh

Validation

• poetry run ruff check articulated/estimation/model.py articulated/rl/agent.py articulated/rl/environment.py
• YAML parse checks for all new RL/estimation configs
• bash -n scripts/train_so2_rnn_hidden_sweep.sh
• End-to-end RL comparison rerun with init-conditioning enabled

Results

New comparison file:

• logs/rl/comparison/obs_mode_comparison_initcond_20260228_014306.csv

Single-seed ranking (higher is better):

• baseline_raw: -3.902
• concat_h128: -8.234
• embed_only_h64: -10.167
• concat_h64: -10.419
• embed_only_h128: -11.054

Notes

Raw baseline still performs best in current runs. Init conditioning improved some embedding variants (notably concat_h128). Multi-seed confirmation is recommended before final conclusions.