Support managed Torch CPU/CUDA runs and live Projects

[MrScripty]

Current Project/model workflows cannot run our CUDA trainer through an installed Torch environment or reliably display a lab project as its files change. This extends the existing workflow with verified CPU/CUDA execution, live Project activity and evidence, and managed-model browser authoring.

Pumas Library continues to own the Python/PyTorch installation. Tuldok registers and verifies that installation, then manages the Run lifecycle; it does not install or upgrade PyTorch.

Changes:

• Add the managed_torch_python_v1 registry, manifest and bootstrap. Bind installed files and symlink targets, select CPU/CUDA explicitly, and preserve existing approvals, cancellation, deadlines, process ownership and artifact publication.
• Support managed Torch bound-model recipes and ordinary managed CPU restore recipes through current Project inspection, capture and browser authoring.
• Add live Project activity, reference/evidence refresh and direct Project links. Show stale/unavailable observations explicitly; cancel obsolete polling and preserve focus and current selections.
• Preserve inspected scientific-policy number representations during browser submission, and avoid false snapshot conflicts when unrelated sibling files are created while still verifying selected files.
• Retain the one-time Rheon lab converter and PostgreSQL fixtures under tools/migrations/rheon-lab-once/, outside application startup, deployed payloads and ordinary CI. Keep the installed-runtime alias regression in normal CI.

Validation:

• Real float32 CUDA training followed by fresh CPU restore, with matching checkpoint/weight and prediction evidence; CPU restore performs zero optimizer updates and does not initialize CUDA.
• Five real lifecycle cases covering cancellation, deadlines, exceptions and cold recovery.
• Unit/contract qualification across runtime, process, model-workflow and snapshot boundaries; actual browser checks cover Project updates, managed authoring and the Model/Study/Trial/Run/checkpoint/pressure-field/CPU-restore chain.
• Final original-catalog browser readback passed in 146.3 seconds, with zero new imports or Runs.
• Offline converter: 16 isolated PostgreSQL tests passed; permanent alias regression passed. Ordinary CI import-only discovery found 1,503 tests, zero loader errors and zero converter imports; this was not a full 1,503-test execution.
• Independent Standards and Spec reviews passed. No model weights, simulation datasets or catalog backups are included.

Operational limits: the installed runtime is trusted local software and must remain unchanged during execution; OS libraries and GPU drivers remain external dependencies. Measured Run-view loading is approximately 28–30 seconds and is recorded separately from model-compute performance.

Commits:

• 824665aefdc68a458b5510a0b33d6cae1eb4887d: managed Torch workflows and live Projects.
• 8af5a6a1e288e5b2f86a75b53cab93ca71222a04: separately retained offline conversion tools and fixtures.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: b462f0ac-e3b7-4de9-812c-9038738e7313

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}