tp2: render only the selected HCA functions into B12X_ROCE_HCA

[SecureBot]

What and why

transport_environment() in the TP2 launcher renders the full four-device hca_inventory into B12X_ROCE_HCA and narrows the selection only in B12X_ROCE_PEER_HCA_MAP / NCCL_IB_HCA. The RoCE proxy opens every device in B12X_ROCE_HCA and requires each port to be ACTIVE, so on a host with one physical DAC in cage p0 the two uncabled f1 functions abort RoCEnante setup and the engine silently falls back to PyNCCL:

WARNING [b12x_roce_all_reduce.py:102] RoCEnante initialization failed: ... RDMA device rocep1s0f1 port 1 is not active
INFO [cuda_communicator.py:314] Using ['PYNCCL'] all-reduce backends ... for group 'tp:0'

This change renders only the selected functions (rocep1s0f0,roceP2p1s0f0) and a peer map addressing that list (0/1). NCCL_IB_HCA, path count, and the profile/manifest hashes are unchanged; profile.json is not touched. With this change RoCEnante initializes as the first all-reduce backend on a single-cable install, matching the README's "both PCI functions of one physical DAC" description.

Fixes #268.

Validation

• python -m pytest runtime/profiles/glm53-flash-spark-tp2 -q — 56 passed (two assertions in test_rank_plan_maps_both_pci_functions_of_one_cage and the rejected site-file peer-map example updated to the new rendering).
• python -m pytest runtime/sparkring/jovian-r33/profiles -q — 33 passed.
• Live: two DGX Spark (GB10), one QSFP112 DAC between port 0 of each, enp1s0f1np1 NO-CARRIER on both, R33 image 3c7779ad71dd… with --r33-sparkcache. Before: warning above on both ranks, ['PYNCCL']. After: no RoCEnante warning; B12X_ROCENANTE selected first for tp:0 and ep:0 on both ranks. FILL IN: any throughput comparison you have (configuration, prompt sizes, units, and that it's a single-site measurement).

I did not re-run the profile's full qualification record; this changes the rendered transport environment, so the maintainers may want to re-qualify or take the alternative of relaxing the ACTIVE requirement in _roce_proxy.c for unmapped devices instead.

Additional context

The tp2-single-dac-source-20260908 record notes both physical cables remained connected during the reference trial, so the single-cable case appears not to have been exercised before. The README transport row is updated to say the other cage may be uncabled.

• I removed credentials and private site identifiers from the change.
• I identified copied or adapted work and updated third-party notices when required. (No copied work.)