Test federated webapp training

[ale18V]

What & why

Part of #1196. Add webapp end-to-end coverage for real federated training with the default Titanic and LUS COVID tasks. Each run drives one browser through the full UI alongside a Node participant and a real Disco server. The tests check participation, training completion, model-sharing progress, and absence of error toasts.

Technical plan

Organize end-to-end training specs under webapp/cypress/e2e/training, with local, federated, and decentralized subdirectories. Ordinary Cypress discovery excludes the dedicated training runs; separate CI jobs select their specs directly. The LUS test uses four epochs and a deterministic, disjoint split of the full image dataset across the two participants. A task-selecting Node peer serves only the model needed by each test.

Deviations from plan

The original Cypress harness and separate config were removed following review; a standalone server/peer process and the existing Cypress config are sufficient. The full LUS dataset is split between participants instead of duplicated on each participant, preserving full-dataset coverage while reducing redundant work. The task-specific package scripts were removed because the CI matrix already selects each federated spec.

Todo

None for this PR. Decentralized and browser-to-browser scenarios remain in stacked #1232 and #1239.

[JulienVig]

The objective of issue #1196 is to test browser-browser collaboration while the current implementation implements browser-Node.js collaboration. For federated learning this doesn't make a big difference since participants don't directly interact with each other but for decentralized learning the current approach doesn't really address the issue. I'm curious if there was a specific reason that pushed you towards relying on Node.js?

Also, I find the current implementation a bit overkill for what it does, do we actually need such a harness? I've pushed a temporary commit with an alternative implementation that doesn't rely on a harness or a separate cypress config (feel free to remove my commit). Were there cases that couldn't not be addressed in a simpler manner?

Let's discuss these points in tomorrow's meeting!

[JulienVig]

Thanks! I noticed that you removed the package.json commands to run the collaborative tests manually, was there a specific reason? I think it would be useful for debugging to be able to run the tests outside of the CI.

Other than that I only have a minor comments, feel free to adapt the suggested comments

[ale18V]

Thanks! I noticed that you removed the package.json commands to run the collaborative tests manually, was there a specific reason? I think it would be useful for debugging to be able to run the tests outside of the CI.

I just wanted to be consistent with the local training PR. Also, the CI has a matrix job, the package.json command was one per task.

I agree it can be useful to have the tests as a command, For each local, federated and decentralized, I can try to add a command testing all tasks maybe

[JulienVig]

Yes could you the commands and document them in contributing.md?

[ale18V]

I looked into it, but it's not as simple as it seems, as we need to start both the peer and the server for the given task. We would need some orchestration to run all tasks with one command.
I added the single command for local training, I will merge this PR and look into it later if we can somehow refactor.
Otherwise an alternative is to just run the command with a default task (e.g titanic)