Add offline uncertainty inference with exact conditioning and joint forecasts

[yichao-liang]

Second of four stacked PRs splitting bridge-learning (commits 2ca1c8453..4fc84c7c4, 56 commits).

Summary

Adds offline uncertainty inference for learned simulators under predicators/code_sim_learning/inference_*.py:

• Legacy inference preserved behind an adapter, with explicit offline replay of continual recordings and observation-channel reconstruction.
• Physical continuation state and candidate prefix replay; Domino task caches keep the initial robot configuration.
• Exact affine/Gaussian conditioning, robot joint and articulated priors, rigid assembly and planar support priors, and rejection sampling of feasible priors.
• Output-error marginalization, coupled orientation error, finger readout reduction, and a complete offline likelihood.
• Joint batch SMC sampling with proposal blocks, tempering schedules, resumable stage checkpoints and process-safe isolated workers.
• Posterior-weighted information scores, assessed joint posterior forecasts, causal future scoring and conditional path integration diagnostics.
• scripts/audit_inference_replay.py plus validation records under docs/uncertainty/.

Checks

yapf, docformatter, isort, mypy and pylint pass locally on this tree.
The full unit test suite was run locally in file shards; the only failures are test_bilevel_sketch_samplers::test_refine_and_validate_report_returns_plan, test_pg3_analogy_approach and test_boil_sketch_refinement, which fail identically on origin/master when run outside CI's shard order.

🤖 Generated with Claude Code