docs: add first scored trace (Notepad Hello World, score 0.5)

6 steps, 91s, GPT-5.4-mini planner+grounder, lightweight mode.
VLM judge passed milestone 2 (Hello World typed, confidence 1.00).
Milestone 1 (process check) timed out during /execute_windows eval.

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

Author: abrichr

docs/traces/notepad_hello_2026_03_21/README.md

@@ -0,0 +1,49 @@
+# Trace: Notepad Hello World — Score 0.5 (1/2 milestones)
+
+**Date**: 2026-03-21
+**Task**: Open Notepad and type "Hello World"
+**Planner**: GPT-5.4-mini (OpenAI)
+**Grounder**: GPT-5.4-mini (OpenAI)
+**Adapter**: WAALiveAdapter (lightweight=True)
+**Steps**: 6
+**Time**: 91.0s
+**Score**: 0.50 (1/2 milestones)
+
+## Milestones
+
+| # | Milestone | Check | Result |
+|---|-----------|-------|--------|
+| 1 | Notepad is open | `Get-Process notepad*` via /execute_windows | FAIL (timeout) |
+| 2 | Hello World typed | VLM screenshot judge | **PASS** (confidence 1.00) |
+
+## Step-by-step
+
+| Step | Action | Screenshot |
+|------|--------|------------|
+| 0 | Reset (clean desktop) | ![step 0](step_00_reset.png) |
+| 1 | Click Start button | ![step 1](step_01.png) |
+| 2 | Start menu open, click Notepad | ![step 2](step_02.png) |
+| 3 | Desktop (Notepad loading) | ![step 3](step_03.png) |
+| 4 | Notepad open, type Hello World | ![step 4](step_04.png) |
+| 5 | Hello World typed | ![step 5](step_05.png) |
+| 6 | Done | ![step 6](step_06.png) |
+
+## What worked
+- Lightweight mode: no cleanup crashes, server stayed responsive
+- GPT-5.4-mini correctly identified Start → Notepad path
+- VLM screenshot evaluation: "PASS (confidence=1.00) — The Notepad window shows the text 'Hello World' clearly in the text area"
+- Task instruction emphasis: planner followed "open Notepad" instead of clicking Chrome
+
+## What didn't work
+- Milestone 1 (process check): PowerShell `Get-Process notepad*` via /execute_windows timed out
+- WAA /evaluate endpoint: unreachable (evaluate_server.py can't connect to Windows VM)
+- OneDrive popup appeared but agent worked around it
+
+## Gap vs customer results (score 1.0)
+Customer scored 1.0 (2/2 milestones) on same task. Differences:
+- They use WAADirect (direct HTTP, no adapter overhead)
+- They skip verify_apps/close_all entirely
+- They use GPT-5.4 (full, not mini) + UI-Venus grounder (not GPT as grounder)
+- Their milestone evaluation runs PowerShell successfully (their /execute_windows is responsive)
+
+Our milestone 1 failed due to /execute_windows timeout during evaluation, not because Notepad wasn't open (screenshot proves it was). The evaluation plumbing is the gap, not the agent behavior.