[codex] compact training sample transport payloads

[samsja]

Summary

Stacked on #2807. This continues the train-trace memory work by compacting the remaining trainer-bound TrainingSample list payloads at the orchestrator/train transport boundary.

What changed

• Adds PackedArray fields to TrainingSample for byte-backed token IDs, masks, logprobs, per-token temperatures, teacher logprobs, and mm_token_type_ids.
• Adds prime_rl.transport.compact helpers to compact samples before send, read compact lengths without inflation, and inflate only when preparing selected trainer microbatches.
• Compacts batch.samples immediately after optional teacher logprobs and immediately before TrainingBatch send.
• Updates trainer prepare_sample to read either legacy lists or packed arrays.
• Updates MultiPacker validation/token-budget accounting to use compact lengths so buffered samples do not inflate just to schedule packing.

Synthetic Size Check

Single synthetic TrainingSample with 30k tokens (15k prompt + 15k completion), scalar completion temperature, no R3 bytes:

• msgpack payload: 315,075 bytes -> 183,882 bytes (58.36% of previous)
• approximate Python list payload for compacted fields: 2,643,088 bytes -> 184,943 bytes (7.00% of previous)

Validation

• uv run pytest tests/unit/orchestrator/test_batch.py tests/unit/train/rl/test_packer_compact.py -> 15 passed
• uv run pytest tests/unit/orchestrator tests/unit/train/rl/test_packer_compact.py -> 89 passed
• uv run ruff check src/prime_rl/transport/types.py src/prime_rl/transport/compact.py src/prime_rl/trainer/batch.py src/prime_rl/trainer/rl/packer.py src/prime_rl/orchestrator/orchestrator.py tests/unit/orchestrator/test_batch.py tests/unit/train/rl/test_packer_compact.py -> passed
• uv run ruff format --check src/prime_rl/transport/types.py src/prime_rl/transport/compact.py src/prime_rl/trainer/batch.py src/prime_rl/trainer/rl/packer.py src/prime_rl/orchestrator/orchestrator.py tests/unit/orchestrator/test_batch.py tests/unit/train/rl/test_packer_compact.py -> passed
• git diff --check -> passed

---

Note

Medium Risk
Changes the hot orchestrator→trainer data path for all training samples; behavior is covered by roundtrip and packer tests and legacy list fields remain supported via accessors.

Overview
Shrinks orchestrator→trainer msgpack traffic and in-process memory by replacing large per-sample Python lists with byte-backed PackedArray fields on TrainingSample, compacting immediately before TrainingBatch send (after optional teacher logprobs).

Adds prime_rl.transport.compact to pack/unpack token IDs, masks, logprobs, temperatures, teacher logprobs, and mm_token_type_ids, expose lengths without inflating lists, and inflate only in prepare_sample when building microbatches. MultiPacker validation and token budgeting use those length helpers so buffered compact samples are not expanded just for scheduling.

^{Reviewed by Cursor Bugbot for commit b0ea1f9. Bugbot is set up for automated code reviews on this repo. Configure here.}