fix(core): validate scan sequence bounds and array dimensions across FTL, latent WM, CBP, and Step 4

[rama0x1]

Summary

Hardens array-based transition scanning and learning loop routines against untrusted array types, incompatible array dimensions/ranks, invalid sequence lengths (num_steps < 1 or overflow), and potential signed-int32 scalar/byte allocation overflows across:

1. alberta_framework.core.ftl_world_model.run_sparse_ftl_world_model
2. alberta_framework.core.latent_world_model.run_latent_world_model_learning_loop
3. alberta_framework.core.continual_backprop.run_cbp_learning_loop
4. alberta_framework.core.option_value_duration.run_option_value_duration_from_arrays
5. alberta_framework.steps.step4.run_step4_scan
6. alberta_framework.core.learners (run_learning_loop, run_learning_loop_batched, run_mlp_learning_loop, run_mlp_learning_loop_batched, run_td_learning_loop, run_true_online_td_loop)

Key Changes

• Exact component/state type verification: Checks model/learner and state objects to prevent silent state mismatches.
• Array metadata enforcement: Verifies input transitions are trusted np.ndarray, jax.Array, or jax.core.Tracer instances with conforming shapes.
• Sequence bounds validation: Ensures 1 <= num_steps <= 2**31 - 1 with canonical int index coercion rejecting boolean/float scalar aliases.
• Resource preflight: Validates that output scalar/byte sizes fit signed int32 bounds prior to scanning.
• Test coverage: Added tests/test_scan_sequence_bounds_validation.py covering positive evaluation and negative invalid type/shape/length error paths.

Validation

• pytest tests/test_scan_sequence_bounds_validation.py -> 6/6 passed
• pytest tests/test_ftl_world_model.py tests/test_latent_world_model.py tests/test_continual_backprop.py tests/test_option_value_duration.py tests/test_step4_production.py tests/test_td_learners.py tests/test_scan_sequence_bounds_validation.py -> 371/371 passed
• ruff check . -> passed
• mypy -> passed

[github-actions]

Automated review ran but produced no report; leaving for a maintainer.

[lalalune]

Blocking: the new validators do not impose a finite scan ceiling. The affected paths call _require_int(..., minimum=1), _require_int32(..., minimum=1), or _require_positive_int(...) without an explicit maximum. Consequently, sequences with hundreds of thousands or millions of steps still pass and can trigger the same pathological JAX trace/compile this PR claims to prevent; the Step 4 path is unbounded as well.

Please use a documented finite ceiling at every affected scan (the sibling convention is 10,000), validate type/rank/leading length before materialization where applicable, and add regression tests that prove an oversized input is rejected before jax.lax.scan/allocation. The companion shape checks are useful, but this is not yet a scan-hang bound.