Memory update from review of PR #4

[arcanist]

Memory updates from PR #4

Add: When adding a new operator using the @operator decorator, especially if the function has a suffix (e.g., _op) that differs from the intended registry name.

Type: gotcha

Memory:

When registering operators with the @operator decorator, always pass an explicit name= parameter. The decorator falls back to func.__name__ when no name is given — so a function named pack_candidates_op registers as "pack_candidates_op", not "pack_candidates". Any code that looks up operators by their intended semantic name (e.g., NoN.from_operators(["pack_candidates"])) will fail silently at lookup time rather than at definition time.

Rationale:

The agent spent significant time tracking down why NoN.from_operators(["pack_candidates", ...]) failed. The bug is non-obvious: the decorator silently uses func.__name__ as the registry key, and the metadata name field is not used. This is a genuine gotcha that a competent engineer would likely miss on first pass.

Why this is a memory (not a convention update): This is a situational pattern that helps the agent in specific contexts, not a permanent rule for all future code.

Referenced files: nons/operators/deterministic.py

---

Add: When designing multi-layer NoN pipelines where one operator returns a list (e.g., List[Candidate]) that feeds into a downstream single-node operator expecting the full collection.

Type: gotcha

Memory:

In Layer._prepare_inputs, list inputs are distributed 1-to-1 across nodes only when len(self.nodes) > 1. For single-node layers, any list input is always broadcast as-is. This means if an operator returns a List[T] and the next layer has exactly one node, that node receives the full list — not individual elements. Be aware of this when chaining operators that return collections through single-node layers.

Rationale:

This was a subtle bug fixed in this session: a single-node layer receiving a 1-element list from extract_winners was unwrapping it and passing a bare Candidate to majority, which expected a list. The fix (guard on len(self.nodes) > 1) is not obvious from reading the code without understanding the intent. Engineers building pipelines need to know this behavior.

Why this is a memory (not a convention update): This is a situational pattern that helps the agent in specific contexts, not a permanent rule for all future code.

Referenced files: nons/core/layer.py

---

Sessions: 5ea2a0a0-6590-4a63-bca5-b2ad4ad1bdbb

Merge to apply. Close to discard.