docs: add model instability debugging page

[gabrielfruet]

What has changed and why?

Adds a new documentation page for Model Instability Debugging (docs/source/debugging/model_instability.md), linked from the main docs nav and from the gradient_norm entry on the Train Settings page.

This is a follow-up to #811 and is intended to be the home for instability-debugging tooling as new methods land. For now it documents gradient norm logging:

• What training instability looks like (loss spikes, NaN/inf collapse, plateaus).
• How to view gradient_norm (console grad_norm, TensorBoard, MLflow, W&B).
• How to interpret the trend (stable / exploding / vanishing) rather than absolute thresholds.
• Common next actions (LR, precision, resume_interrupted, normalization).

The page is structured so future debugging tools can be appended under the same section.

How has it been tested?

• mdformat --check on all three changed files → exit 0.
• sphinx-build -b html --fail-on-warning --keep-going source build/local → exit 0, no warnings.
  Built with PYTHONPATH=../src ../.venv/bin/sphinx-build (worktree venv editable-source gotcha).
• Verified both cross-references (../settings/train_settings.md, ../faq.md) resolve.

Did you update CHANGELOG.md?

• Yes
• Not needed (documentation-only change)

Did you update the documentation?

• Yes
• Not needed (internal change without effects for user)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 09843e56bf

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Don't resume interrupted runs after changing settings

In the mitigation flow where the user lowers LR or changes precision, this tells them to use resume_interrupted after changing settings. The train settings page documents resume_interrupted as crash recovery only and says training arguments must not be modified; the training loop restores the optimizer and scheduler state from out/checkpoints/last.ckpt, so changed hyperparameters like LR/schedules may be overwritten or mixed with the old run state. Users following this advice may not actually test the intended fix and can keep resuming the unstable run; this should point to the “load checkpoint for a new run” flow instead.

Useful? React with 👍 / 👎.

[gabrielfruet]

/review

[gabrielfruet]

/review

[mrpositron]

LGTM!