docs(verify): record that GNN promotion breaks API serving at startup (BR-1)

[Gavin-Borges]

Rescues BR-1, an operational warning that existed on exactly one local branch, on no remote, and on no task list. Docstring only; no behaviour change.

What it records

Passing the five GNN promotion gates is necessary but not sufficient to ship a GNN, and nothing in the repository said so.

promote_model rewrites config.yaml's model_path to the GNN checkpoint without relocating it. But ModelManager.load passes only config.model_path.name - the basename - to ModelRegistry.load, and resolve_model rebuilds that as models/<basename> relative to cwd. The directory component is discarded, the lookup misses, and FileNotFoundError is raised inside the unguarded startup lifespan handler.

So promoting a GNN makes the API fail to start, /health included. It does not degrade, and it does not fall back.

Two further defects sit latent behind that one: a .pth loads as a raw state_dict with no predict_proba, and _score_peptide builds a flat 31-float vector where GraphPredictorV2.forward expects a PyG batch. There is no model-type dispatch in api/main.py at all.

Re-verified against current origin/main line by line before this branch was cut, not carried forward on the original commit's word: api/main.py:204 passes self.config.model_path.name; model_registry.py:14-19 rebuilds it as models/<basename> and raises only on directory escape; model_registry.py:52 raises the FileNotFoundError; lifespan at api/main.py:342 is _manager.load() with no try/except.

Why it is a fresh branch rather than the original commit

The original commit 27ac579 is authored and signed off as Test <test@example.com>, a placeholder identity from a git misconfiguration - neither correct nor DCO-valid on a repository with a single author of record. It could not be pushed as-is.

This is a clean cherry-pick onto current main, re-authored to Gavin Borges with a matching DCO trailer and GPG-signed. Content is byte-identical to the original; the cherry-pick applied with no conflicts. The provenance is recorded in the commit message rather than lost.

27ac579 is the twentieth commit of the same Test <test@example.com> window. The other nineteen are on published history and are handled by .mailmap in #288, which maps display without rewriting any commit object. This one was never pushed, so it could be re-authored properly instead - which is strictly better, and is why it was deliberately excluded from that mapping.