aetheric-gpu 0.1.0-alpha

# Stability Test — Sustained Inference Pressure

Goal: validate that the Llama-7B benchmark can run continuously for 6+ hours
without memory leaks, NUMA drift, thread pool starvation, or compressed-RAM
backend transitions degrading throughput below the 1.6x envelope.

## Methodology
- Run the existing `llama_bench.rs` in a loop, capturing per-cycle token/sec.
- Each cycle = full Llama-7B forward pass (cache stripped between cycles).
- Every N cycles, log: RSS (Resident Set Size), token/sec, p99 latency.
- Pass criteria: token/sec stays within 1.42x envelope (>= 12.0 tok/s)
                and RSS does not grow > 30% over baseline after 6h.

## Running
```bash
while true; do
  ./target/release/examples/llama_bench --no-warmup 2>&1 | tee -a stability.log
  date -u +"%Y-%m-%dT%H:%M:%SZ" | xargs -I{} echo "[{}] cycle complete" | tee -a stability.log
done
```

## Acceptance criteria (honest)
- [ ] 1h sustained run: token/sec degrades < 10% from initial
- [ ] 4h sustained run: RSS growth < 20%
- [ ] 6h sustained run: token/sec still within 1.6x dGPU envelope (11.94 baseline)
- [ ] No panics, mutex deadlocks, or thread starvation