Release v0.34.33 - "Facsimile"

Release Date: August 15, 2026

Named for what this release found: files that were present, committed, and unreadable. Model metadata had been added to version control back in February, but stored as Git LFS pointers — so any checkout without LFS got a stand-in for the file rather than the file, and the machine-learning code quietly fell back to defaults instead of failing.

Bug Fixes

• Model metadata that live loaders require is now committed MBA-1449 — a repo-wide *.json ignore rule was excluding metadata that sits beside tracked model artifacts, so a fresh clone received a model without the information needed to run it. Seven files are now tracked, each verified load-bearing by finding the code that reads it: two are required for the model to load at all, and five supply the feature names used to order the inputs handed to the scaler — without which predictions are silently computed on mis-ordered columns rather than raising an error. Training reports in the same directories have no runtime reader and remain untracked.
• Load-bearing metadata is no longer stored as Git LFS pointers MBA-1449 — tracking the files turned out not to be enough. They were being routed into LFS, so a checkout without LFS wrote pointer text to those paths; parsing then failed, and every loader swallowed the error. This meant an earlier fix from February had never worked in the situation it was written for: the file it addressed was committed, but committed as a pointer. Twelve files now store real content, verified by checking out the repository with LFS deliberately disabled — what a continuous-integration runner without LFS sees.

Documentation

• Recorded that Lambda subnets must be added to the fck-nat security group — a route to the NAT instance is not sufficient for outbound traffic; the instance's own security group lists permitted subnets explicitly, and a subnet that is routed but unlisted has its traffic dropped rather than rejected, so connections hang for the full 30-second function timeout instead of failing fast.

Testing

• 26 new tests assert the pairing — a tracked model artifact must not have untracked metadata — that each file carries the keys its loader actually reads, and that the committed blob parses as JSON rather than trusting the working copy. The working copy looked correct throughout, which is exactly why the earlier fix appeared to succeed.

← All Releases