A committed adaptive model with 41.5% training accuracy loads automatically when the server starts from v2/ #2135

v2/data/adaptive_model.json is tracked in the repository. It records trained_frames: 3316, training_accuracy: 0.4149577804583836 and version: 1, which is 41.5% accuracy, measured on the training data itself.

The server loads it at startup from a relative path. adaptive_classifier::model_path() returns data/adaptive_model.json (v2/crates/wifi-densepose-sensing-server/src/adaptive_classifier.rs, around line 675), and the AppState initialiser in main.rs (around line 12297) calls AdaptiveModel::load on it. Any server started with v2/ as the working directory therefore logs Loaded adaptive classifier: 3316 frames, 41.5% accuracy and applies the model as an override to the per-node classification label (the "Adaptive override" block in the raw-CSI path of main.rs). A server started from anywhere else doesn't find it.

CODE-DERIVED from reading the code, plus the log line, which I saw in test runs started from v2/. I haven't measured how its labels compare with the built-in classifier on real data.

Two things follow.

• Behaviour depends on the directory the server is started from, and on a model nobody on a fresh checkout trained.
• Because it is version: 1, it also keeps the legacy subcarrier-proxy for dominant frequency instead of the measured temporal frequency used by version 2 models.

The file is 6211 bytes and first appears in this history in 82c1b8f.

Should this file be removed from the repository, or stay and be loaded only with an explicit flag? If it's meant as a sample, moving it out of the default load path, for example under examples/ or docs/, would keep the behaviour from changing silently. I can send a PR for whichever you prefer.