Add AstraPush random HPT pilot foundation and OSMO preflight

[rpuns]

Implements the foundation of the supplied AstraPush pilot inside EgoVerse-graph: a fully random HPT graph, byte-language stem, masked flow targets, typed LIBERO scene compilation, independent task semantics, immutable HDF5 data and resumable episode-balanced sampling. The declared policy predicts ten 10 Hz actions and executes one before replanning.

Stacks on #159 to reuse the consolidated HPT graph. The original version 2 specification is preserved verbatim alongside an implementation plan. Existing experiments and branches are preserved.

Validation: 49 focused tests and 13 configuration/inference checks passed, including recursive composition/construction of all shipped YAMLs. Full offline construction has 45,934,983 trainable parameters and no weight-loading/network calls. S1 and S3 custom scenes compiled and rendered locally in both cameras; previews and receipts are included.

Draft: the source-pinned one-L40 OSMO preflight verifies actual Astra access, a synthetic engineering update/inference and the separate locked simulator environment. It does not claim a real-data W01 pass or production training. Teacher commissioning, the six-scene novelty gate, complete round orchestration and paired learner evaluation remain follow-up work. See docs/experiments/astra-hpt-libero/IMPLEMENTATION.md.