Release unused CLI checkpoint payload before training

[CompleteDotTech]

Change

Refs #3.

Release the CLI checkpoint dictionary before entering train. The model and tokenizer remain owned by training; the initialization parent digest is saved separately, and resume still supplies no initialization parent. This avoids retaining unused deserialized weights and optimizer tensors while the trainer loads resume state again.

Four weak-reference regressions cover train/reward-train and init/resume. They assert payload collection, two-step configuration and unchanged lineage/resume arguments. Earlier source fails all four; the fix passes all four. This establishes object ownership; actual full-checkpoint RSS savings and native full-weight feasibility remain unmeasured.

Validation

• Full CPU suite: 862 passed, 55 explicitly skipped, 63.60 seconds. CUDA/native, SDK and platform skips remain explicit.
• Compilation, indexed source audit228files, critical Ruff0.13.2, feature-matrix identity, lock and diff checks: PASS.
• CPU24-configuration numerical audit and local same-policy resume/calibration exercise: PASS, fixture scope.
• Independent code/regression review: no defects found.

No optimizer math, configuration identity, telemetry or API shape changes.