feat(miles): allow tms preload on Blackwell with cu13+ torch wheels

[zhenyulincs]

What

Make assert_tms_hook_mode_matches_arch CUDA-version-aware: preload hook mode on Blackwell (cc major >= 10) is now allowed when torch.version.cuda parses to major >= 13. Pre-cu13 / unknown CUDA stays blocked (RuntimeError with updated remedies); the MILES_TMS_ALLOW_PRELOAD_ON_BLACKWELL=1 escape hatch is unchanged.

Also fixes env plumbing so the knobs actually reach the actor process:

• both rlix example drivers now forward MILES_TMS_ALLOW_PRELOAD_ON_BLACKWELL + MILES_MAX_RESIDUAL_GPU_MEM_GB
• both actor_group.py spawn paths now forward MILES_TMS_HOOK_MODE + the escape hatch (the legacy colocate path forwarded neither; the guard reads them inside the actor)
• stale "CUDA 12.9" comments refreshed

Why

2026-07-05 offload audit (4x RTX 5090, torch 2.11.0+cu130): the historical tms 0.0.9 preload segfault is gone on cu13 wheels — tms mock pause/resume PASSES and the full dual-pipeline E2E smoke completes with EXIT_CODE=0 in preload mode. Preload is the only hook mode whose catchment includes Megatron optimizer state (torch-hook leaves ~11.07 GiB whole-GPU residual at cycle 2 vs 5.02 GiB stable under preload), which unblocks lowering the rlix residual gate from 13.0 to 7.0 (paired rlix PR).

The guard's dev-note anticipated exactly this follow-up: relax once a fixed tms/CUDA build is confirmed.

Tests

PYTHONPATH=. python -m pytest tests/fast/backends/test_tms_utils.py tests/fast/utils/test_rlix_env_forwarding.py -q
32 passed  (on 4x RTX 5090 cu130; torch.version.cuda == '13.0' confirmed)

New coverage: cu13.0/13.1 allowed (env preload AND env-unset default), cu12.9/None/unparseable raise, escape hatch on cu12.9, pre-Blackwell unchanged; source-level tests pin the four env-forwarding requirements.

E2E

Paired trio (this + #5 + the rlix threshold-7.0 PR) dual smoke on 4x RTX 5090: EXIT_CODE=0, whole-GPU gate readings 3.45-3.47 GB (cycle 1) / 5.02-5.04 GB (cycle 2) vs threshold 7.00 GB.

Codex review: APPROVE_WITH_NOTES, 0 blockers.

Refs: rlops/rlix#17 (merged; audit evidence in its description), #5 (paired sender-side diagnostics).

[zhenyulincs]

Closing per maintainer decision: no miles-side code change needed. The tms arch guard already ships an escape hatch (MILES_TMS_ALLOW_PRELOAD_ON_BLACKWELL=1) that authorizes preload on stacks where the cu12.9 segfault is confirmed gone; the rlix smoke (rlops/rlix#31) now exports it directly. The audit evidence (preload works on cu130 Blackwell, residual 5.02 GiB) stands and is recorded in rlops/rlix#17 / #31.