A trained model outlives its process, so the learned half of the loop closes too

[syncytium2]

The settings half landed in #507; this is the other one. docs/pipeline.md's blocker list
is now empty — both items that sat upstream of everything else on that page are closed.

• bugarach.learn.checkpoint saves and loads a trained model.
• run_learned_on_folder.py --save-models writes them.
• bugarach detect --model ckpt.json runs one over an export folder, repeatable, beside
  the six.

Verified the only way that counts

On the pilot APV+CNQX+GZ folder: train two models and save them, then detect with those files
from a separate process.

model / stream	in-process	from file
tube / fast	101	101
tube / slow	253	253
tube_guard / fast	112	112
tube_guard / slow	257	257

Identical call for call. 31 KB per checkpoint.

JSON, not torch.save

These nets are 1,149–2,393 parameters — smallness is their claimed advantage — so a
checkpoint is tens of kilobytes of numbers and can be a format a person reads, a diff shows
and JSON.parse loads. That is what makes a model shareable in both directions rather
than only out of one.

And torch.save is pickle. ADR-0005's target flow is a user downloading a folder of models
from the site; a format where opening a stranger's model executes their code cannot be that
format. Nothing here needs it — the state dict is arrays of floats.

What travels with the weights

The architecture name and the config it was built with, so moving a registry default
cannot rebuild a different network under old weights. The encoding contract — dt, the
onset field, the busiest-first row rule — because a model is a function of the raster it was
shown, and encode's own docstring is emphatic that row order is a coordinate rather than a
label. The operating point, since a model without its threshold is not a detector. And
the provenance, including the torch thread count, because train.THREADS records
reduction order moving F1 by 0.018 on this very code.

What it refuses

A tensor whose shape no longer matches, with both shapes named — that is the failure a
saved model actually has, the file outliving an edit to the architecture, and a half-loaded
model that then scores is worse than one that will not open. A changed encoding contract, for
the same reason. A model whose name collides with one of the six, because detections.csv's
detector column is one namespace across both families. And a bad checkpoint fails on
recording 0, not on 84 of 85.

peek reads provenance without importing torch, so a listing need not instantiate a
network to say what a model is.

Learned calls land in the six ports' contract and are scored inside each analysis window,
the way the flat three are, so no model gets context across a drug transition the others were
denied. Being indistinguishable from a hand-written call in the file is deliberate — which is
why run.json now carries the roster of models that ran. Their n_roi stays NA: the model
emits a per-frame score and never says which cells it meant.

Full suite 2260 passed, 33 skipped, 1 xfailed. Sapper clear.

🤖 Generated with Claude Code