Fix AMD GPU power measurement with rocm-smi (#4)

## Problem
The rocm-smi energy meter was returning 0.0 for all power readings on
AMD GPUs.

## Root Cause
1. Used `-p` flag which shows 'Performance Level', not power consumption
2. Regex patterns didn't match the actual output format from modern
rocm-smi

## Fix
- Changed flag from `-p` to `--showpower`
- Added regex pattern for `Current Socket Graphics Package Power (W):
X.XXX` format

## Testing
Verified on AMD Radeon Graphics (RYZEN AI MAX+ 395):
```
Before: energy_wh_raw: 0.0
After:  energy_wh_raw: 0.0022 (2 second test)
```

Also includes copilot-instructions.md for AI coding agents.

Author: NOVADEDOG

.github/copilot-instructions.md

@@ -0,0 +1,98 @@
+# Copilot Instructions for Energy Leaderboard Runner
+
+## Project Overview
+
+This project measures real-world energy consumption of local LLMs. It has two main components:
+1. **Python CLI** (`src/`) - Runs benchmarks and collects energy metrics from hardware sensors
+2. **React Web App** (`energy-leaderboard-web/`) - Displays crowdsourced benchmark results
+
+## Architecture & Key Patterns
+
+### Energy Meter System (Plugin Architecture)
+All energy meters inherit from `EnergyMeter` base class in [src/energy_meter/base.py](src/energy_meter/base.py):
+- `is_available()` → Platform detection
+- `start()` / `stop()` → Measurement lifecycle returning `energy_wh_raw`, `duration_s`, `sampling_ms`
+
+Platform detection priority in [src/energy_meter/integrator.py](src/energy_meter/integrator.py):
+- macOS: `powermetrics` (requires sudo)
+- Linux: NVML (NVIDIA) → ROCm (AMD) → RAPL (CPU)
+
+### LLM Integrations (Plugin Architecture)
+Implement `LlmRunner` from [src/llm_integrations/base.py](src/llm_integrations/base.py):
+- `check_connection()` → Validate endpoint
+- `generate()` → Returns `(text, tokens_prompt, tokens_completion, response_time_s)`
+
+Supported: Ollama (`ollama_client.py`), OpenAI-compatible (`openai_client.py`)
+
+### Configuration
+Environment-based via [src/config.py](src/config.py):
+- `OLLAMA_HOST` (default: `http://localhost:11434`)
+- `CO2_INTENSITY_G_KWH` (default: 350.0 - EU average)
+- `SAMPLING_INTERVAL_MS` (default: 100)
+
+## Commands & Workflows
+
+```bash
+# Install dependencies
+pip install -r requirements.txt
+
+# Run single benchmark (requires Ollama running)
+python src/main.py run-test --model llama3:latest --test-set easy
+
+# Run all test sets (easy, medium, hard, mixed) - preferred for contributions
+python run_all_tests.py --model llama3:latest
+
+# OpenAI-compatible provider
+python run_all_tests.py --model gpt-4 --provider openai --base-url https://api.example.com
+
+# Docker (Linux with GPU)
+docker build -t energy-leaderboard-runner .
+docker run --rm --gpus all -v $(pwd)/results:/app/results \
+  -e OLLAMA_HOST=http://172.17.0.1:11434 energy-leaderboard-runner \
+  run_all_tests.py --model llama3:latest
+```
+
+## Test Sets
+
+Located in [src/data/testsets/](src/data/testsets/). Structure:
+```json
+{
+  "id": "ts1",
+  "name": "...",
+  "goal": "...",
+  "questions": [{ "id": "...", "prompt": "...", "difficulty": "easy" }]
+}
+```
+Reference testset by name without prefix: `--test-set easy` (resolves to `testset_easy.json`)
+
+## Output Schema
+
+Results validated against [src/data/metrics_schema.json](src/data/metrics_schema.json). Key metrics:
+- `energy_wh_raw` / `energy_wh_net` - Energy consumption
+- `wh_per_1k_tokens` - Efficiency metric
+- `g_co2` - Calculated from `CO2_INTENSITY_G_KWH`
+
+Output files: `results/output_{model}_{testset}_{date}.json`
+
+## Web Frontend (`energy-leaderboard-web/`)
+
+React + Vite + Tailwind. Data lives in `public/data/*.json`.
+
+```bash
+cd energy-leaderboard-web
+npm install && npm run dev    # Development
+npm run build                 # Production build
+```
+
+## Contributing Benchmarks
+
+1. Run `python run_all_tests.py --model <model>`
+2. Copy `results/output_*.json` → `energy-leaderboard-web/public/data/`
+3. Submit PR with new JSON files
+
+## Code Conventions
+
+- CLI uses Typer with Rich for console output
+- Dual import pattern in main.py supports both module and direct execution
+- Abstract base classes define interfaces; implementations are in same directory
+- All file paths use `pathlib.Path`

src/energy_meter/rocm_smi_meter.py

@@ -73,9 +73,9 @@ def _poll_power(self) -> None:
             try:
                 # Run rocm-smi to get power usage
                 # -d: device index
-                # -p: show power consumption
+                # --showpower: show current power consumption
                 result = subprocess.run(
-                    ["rocm-smi", "-d", str(self.device_index), "-p"],
+                    ["rocm-smi", "-d", str(self.device_index), "--showpower"],
                     capture_output=True,
                     text=True,
                     timeout=1.0,
@@ -102,9 +102,12 @@ def _parse_power(self, output: str) -> Optional[float]:
         Returns:
             Power in watts, or None if parsing fails.
         """
-        # Look for patterns like "Average Graphics Package Power: 123.45 W"
-        # or "Power: 123 W"
+        # Look for patterns like:
+        # "Current Socket Graphics Package Power (W): 4.052"
+        # "Average Graphics Package Power: 123.45 W"
+        # "Power: 123 W"
         patterns = [
+            r"Current Socket Graphics Package Power \(W\):\s+([\d.]+)",
             r"Average Graphics Package Power:\s+([\d.]+)\s*W",
             r"Power:\s+([\d.]+)\s*W",
             r"GPU Power:\s+([\d.]+)\s*W",