feat(api): extend /api/health with model load diagnostics

[Swatikantamishra8]

Summary

Closes #20

Extends the /api/health endpoint to report per-model load status, addressing the feature request in issue #20.

Changes

• Added model_diagnostics dict to /api/health response
• For each enabled analysis type, attempts to load the model via _load_model()
• Reports loaded (bool), path (checkpoint path), and error (if any) per model
• Health status is marked as degraded if any model fails to load

Example Response

{
  "status": "ok",
  "version": "0.2.0",
  "analysis_types": ["deforestation", "ice_melt", "flooding"],
  "config_valid": true,
  "config_issues": [],
  "model_diagnostics": {
    "deforestation": {"loaded": true, "path": "models/best_model.pth", "error": null},
    "ice_melt": {"loaded": false, "path": null, "error": "No checkpoint found"}
  }
}

Notes

• model_diagnostics values are dict[str, Any] with keys: loaded, path, error
• Models that load successfully will have error: null
• This is non-breaking: existing fields are unchanged

[femi23]

Thanks for the contribution @Swatikantamishra8 — the intent (surfacing per-model load status from /api/health) is exactly right and matches what we need for the Kubernetes readiness probe story. Unfortunately the patch as it stands won't parse, so a re-push is needed before we can land it.

Blockers

1. The diff has trailing whitespace on the module docstring from __future__ import annotations line. Minor, but combined with #2 it suggests the patch went through an editor that mangled whitespace.

2. The added block is not valid Python — indentation is broken. In health() the diff inserts:

          model_diagnostics: dict[str, Any] = {}
          from climatevision.inference.pipeline import _load_model, _find_best_checkpoint
          for atype in enabled_types:
                    name = atype["name"]
                    mstatus: dict[str, Any] = {"loaded": False, "path": None, "error": None}
                    try:
                                  _load_model(name)
                                  mp = _find_best_checkpoint(name)
                                  mstatus["loaded"] = True
                                  mstatus["path"] = str(mp) if mp else None
                              except Exception as exc:
                                            mstatus["error"] = str(exc)
                                        model_diagnostics[name] = mstatus

Three problems:

• The leading indent is 10 spaces; the surrounding function body uses 8.
• try: body is indented 30 spaces, but except is at 30 too and offset by more spaces than try — Python will raise IndentationError.
• The final model_diagnostics[name] = mstatus sits inside the except block instead of after it, so the success path never populates the dict.

Please run python -m py_compile src/climatevision/api/main.py locally before pushing — that'll catch this in one shot. CI should fail this too, which makes me think the diff wasn't actually pushed/tested locally before being sent.

3. _load_model and _find_best_checkpoint are module-private (leading underscore). Reaching into private helpers from another module breaks our pipeline.py contract. Two cleaner options:

• Expose a public get_model_load_status(name) -> dict from inference/pipeline.py (preferred — keeps the diagnostics shape behind one API).
• Or move the diagnostics logic into pipeline.py and call a single public function from health().

Should-fix

4. Calling _load_model for every analysis type on every /health hit is expensive if the model isn't already cached — it does disk I/O and potentially a torch state-dict load. The health endpoint is hit by load balancers every few seconds in prod, so this could DoS our own checkpoints folder. Please:

• read from the existing in-memory cache only (don't force-load)
• or guard the deep diagnostics behind a ?deep=true query param and have the default response stay cheap (just cached: true/false).

5. Please add a test. Something like tests/test_health.py::test_health_includes_model_diagnostics asserting the response contains model_diagnostics: {<name>: {loaded, path, error}} for each enabled analysis type, with _load_model patched to (a) succeed and (b) raise.

Once the syntax is fixed and we have a public accessor + caching guard, I think the rest will be quick. Looking forward to v2.

[Goldokpa]

📢 Heads-up: repo history was rewritten today (2026-05-18)

We force-pushed a cleaned history across all branches to remove an internal directory from past commits. Your code and this PR are unaffected — only the commit SHAs underneath have shifted. GitHub will re-render the diff against the new base automatically.

If you have a local clone, please bring it back in sync before pushing anything else:

# Option A (simplest): fresh start
git clone https://github.com/Climate-Vision/ClimateVision.git

# Option B: rebase the existing PR branch in your fork
git fetch origin
git checkout <your-branch>
git rebase origin/main          # likely no conflicts
git push --force-with-lease

Do not git pull on an existing clone — it will produce a messy non-fast-forward state. Either re-clone, or rebase explicitly as above.

Apologies for the interruption — really appreciate your patience here. If anything looks off after rebasing, leave a comment and I'll help unblock right away. Thanks for contributing 🙏