Add feedback loop for cross-project pattern registry

- Create scripts/utils/propose-pattern.js to propose new patterns or
  extend existing ones based on a newly applied workaround/fix.
- Default to dry-run preview; --apply writes files only after human review.
- Matches against existing patterns via keyword overlap and proposes an
  update when the shape is similar enough, or a new pattern file otherwise.
- Update patterns/README.md with feedback-loop instructions.
- Update MEMORY.md to document the plasticity loop.
- Update CONTEXT.md decisions log and current tasks.

MEMORY.md

@@ -75,6 +75,8 @@ See:
 
 Patterns are automatically loaded by `architecture/pipeline.js` when the task input matches a pattern's tags or affected projects. This is the generalization layer of the agent continual-learning system.
 
+**Feedback loop:** when an agent applies a reusable workaround or fix, it should run `scripts/utils/propose-pattern.js` to propose a new pattern or extend an existing one. The helper is dry-run by default; use `--apply` only after human review. This is the plasticity layer — the registry grows from real work instead of being reconstructed from memory later.
+
 Established: 2026-08-06.
 
 ### Small Local Ollama Models and Malformed JSON