chore: clean up superproject structure

Move content to correct locations:
- AGENTS.md → .agents/AGENTS.md (with orchestrator reference update)
- third_party/docs/ (8 reference docs) → wiki/concepts/
- manuals/ (12 Node-RED docs) → wiki/manuals/

Delete 23 unreferenced one-off scripts from scripts/ (keeping 5 active).
Delete stale Dockerfile.e2e, docker-compose.e2e.yml, test/e2e/.
Remove empty third_party/ directory.

Root is now: README, CLAUDE.md, LICENSE, package.json, Makefile,
Dockerfile, docker-compose.yml, docker/, scripts/ (5), nodes/, wiki/,
plus dotfiles (.agents, .claude, .gitea).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

wiki/manuals/node-red/runtime-node-js.md

@@ -0,0 +1,31 @@
+# Node-RED Runtime Node JS Manual (EVOLV Summary)
+
+Based on: https://nodered.org/docs/creating-nodes/node-js
+
+## Input Handler Contract
+
+- Node-RED calls runtime handlers as `function(msg, send, done)`.
+- For compatibility, custom nodes should support fallback behavior when `send` is not provided.
+- Use `done()` on successful completion and `done(err)` on failure.
+
+## Output Routing Rules
+
+- For single-output nodes: `send(msg)` sends on output 1.
+- For multi-output nodes: pass an array with one slot per output, for example:
+  - `[msg, null, null]` => output 1 only
+  - `[null, msg, null]` => output 2 only
+- Always construct the outbound message first, then send it.
+
+## EVOLV Runtime Pattern
+
+Recommended runtime input structure for `src/nodeClass.js`:
+
+1. Normalize `send` fallback.
+2. Wrap `switch(msg.topic)` in `try/catch`.
+3. Route response messages to explicit output index with array format.
+4. Call `done()`/`done(err)` consistently.
+
+## Common Failure Mode
+
+- Sending a bare object in a multi-output context can hide intended port routing.
+- In EVOLV nodes with 3 outputs (`process`, `dbase`, `parent`), use explicit arrays for deterministic wiring behavior.