[Jake Smith (jsmith13)]

Summary

• upgrade vulnerable protobuf pins to 5.29.6, fixing CVE-2026-0994 and CVE-2025-4565
• upgrade the coupled TensorBoard compatibility set to tensorboard==2.12.0 and tensorboardX==2.6.1
• prevent unpinned Conda TensorBoard installs from replacing the secure resolution
• constrain legacy PyTorch 1.9 install paths to numpy<2

A protobuf-only bump, as proposed by #219, does not resolve with TensorBoard 2.11.2 and breaks TensorboardX 2.4.1 at runtime. This updates the compatibility set together without changing Fairseq, PyTorch, model code, checkpoints, or training settings.

Validation

• built clean pre-patch and patched Docker images successfully
• docker/verify_environment.py and pip check pass in both images
• loaded historical Graphormer redox checkpoint d5c5cb62cf17217eadfb5b334e568ced000e0f6eb5871ce81bd62851a4853f09
• compared 20 inference outputs across batched, repeated, reversed-order, and singleton calls: 20/20 exact matches, maximum absolute difference 0.0
• ran two deterministic fine-tuning epochs in both images: losses and gradient norms matched exactly at all four optimizer steps
• final fine-tuned model hashes matched exactly; post-training predictions matched 8/8 with maximum absolute difference 0.0
• TensorBoard event writing succeeded during fine-tuning with protobuf 5.29.6

[gabrieldernbach (gabrieldernbach)]

lgtm