fix: enforce one-server-at-a-time workflow in agent protocol

Agents were attempting to build/validate multiple MCP servers in
parallel, leading to incomplete work that never finished. This adds
a mandatory rule: work on exactly one MCP/LSP server end-to-end
(build, handshake, live client invocation with real data, status
update, commit) before starting the next.

💘 Generated with Crush

Assisted-by: Crush:glm-5.2

AGENTS.md

@@ -159,6 +159,27 @@ git config core.hooksPath .githooks
 
 If you cannot complete an item (e.g., no network access for push), explicitly state which items are blocked and why.
 
+## One Server At A Time (MANDATORY)
+
+AI agents MUST work on exactly **one MCP/LSP server at a time, end-to-end** before
+starting the next one. This includes:
+
+1. **Build** the container
+2. **Run** the container and validate the protocol handshake
+3. **Spin up a live Crush or bash session** and actually invoke the server through
+   the real client pipeline
+4. **Observe a successful real tool call** (not just a handshake — an actual tool
+   invocation that returns real data or completes a real action)
+5. **Update STATUS.md and JOURNAL.md** for that server
+6. **Commit and push** before moving on
+
+**NEVER** attempt to build/validate multiple MCP servers in parallel or batch them.
+**NEVER** skip from "built" to the next server without completing steps 3-4 (real
+client invocation with real data).
+
+The pattern of "build all, then validate all" leads to incomplete work that never
+finishes. One server, fully done, then the next.
+
 ## Related Documentation
 
 - [README.md](README.md) - Project overview, server inventory, installation, usage