[pzhu-flexiv]

Labels are given on the Labelling Frames page (#60), but the learned check could only be trained from the command line. This adds Train this check to the same page, so step 6 of Stage 1 (label, then train the "latched" check) needs no terminal.

What changes

• Panel: under the label bar:
  • Inputs: cameras (from the loaded datasets, each with an optional crop x0,y0,x1,y1), state columns, and the folder to save to (default: the label's name).
  • Train runs on the labels of every loaded dataset.
  • While it runs: elapsed time and the last log line (epoch 12/30: loss …), with Stop.
  • When it ends: the held-out report, false passes first, then each held-out episode's result (false passes and misses highlighted), and how a plan names the check (checks/<name>). A failure shows the command's own error: line.
  • It refuses a bad crop before sending anything.
• Job: it runs the same command the README gives (python -m flexivtrainer.evaluation.learned_check train …) in its own process: one at a time, so the server stays responsive and the GPU memory is freed when it ends.
  • POST /checks/train starts it. Datasets are held to the storage root (403/404). A bad request gets 400, and so does a name already taken. One already running gets 409.
  • GET /checks/train gives its state, log tail and result. POST /checks/train/stop stops it.
  • GET /checks lists the trained checks under <storage>/checks/ with what they look at and their held-out reports.
• Label view: /datasets/check-labels also returns each dataset's cameras (depth maps left out) and state columns, for the form.
• README: the learned-check section explains the panel.

Tests

• Backend: tests/test_check_training.py covers:

  • the exact command built from a request, and 6 refusals;
  • a real training run in its own process (pixels encoder, 3 epochs, on a labelled LeRobot dataset), its held-out report, and GET /checks listing it;
  • a failing command reporting its error: line;
  • a running job refusing a second one, then stopping;
  • routes keeping datasets inside the storage root.

• Panel logic: tests/label_frames_ui.test.cjs gains 3 tests:

  • the camera and state choices;
  • the request it sends;
  • Train enabled only with an input chosen;
  • a bad crop refused before sending;
  • the running, done and failed views.

  Each of 5 mutations is caught.

• By hand: headless Firefox against a test server on a scratch copy of a recorded dataset (4 episodes labelled through the API). I chose wrist_1 with a crop and Fz, then pressed Train. It trained ResNet-18 on the CPU in its own process, then showed the held-out report and saved checks/seated_ui. The report itself means nothing, as the labels were arbitrary. No page errors.

Note for merging

This PR and #61/#62 each bump the static asset version (?v=) in index.html and three tests. Whichever merges second has that one conflict; I'll resolve it.

Review fixes (second commit)

An independent review reproduced two real races in the job. Both are fixed and tested.

• Start while stopping: a new job could start while the last was stopping. The old job's thread then took over the new one: it marked it failed and left its process untracked, so a second GPU training could start. Now starting is refused while a job runs or stops, and each process's threads only touch their own job.
• Stop racing the end: a stop that raced the end reported "stopped" for a check that had been saved. Now:
  • an exit with the result counts as done;
  • a check saved before the stop is kept, and the status says so (saved);
  • a check folder left half-written by a stop or a failure mid-save is removed, so it no longer blocks its name;
  • stop no longer signals a process that has already ended.
• Page:
  • The Stop button was rebuilt every second and could lose a click; only the progress line changes now.
  • A job found running when the page opens now locks the form at once.
• Smaller:
  • crops keep every digit on the way to the command;
  • names must match whole (a trailing newline was accepted);
  • GET /checks reads the folder training writes to;
  • the server stops a training when it stops;
  • the child process speaks UTF-8 on Windows;
  • only video cameras and listed state names are offered;
  • escapeHtml escapes quotes. This is the same change as Chains: judge each run, and count clean runs in a row #62, in identical text, so the two merge cleanly.