stability(queen): long-term training stability study (10000+ episodes)

[gHashTag]

Summary

Training stability monitor for long-term Queen training runs (10,000+ episodes).

New file

• src/b2t/stability_monitor.zig — 223 LOC

Features

• Per-episode metrics: loss, reward, success, memory, CPU, elapsed
• Divergence detection: loss spike > threshold triggers count
• Memory budget checking: alert on overrun
• Log every 100 episodes, checkpoint every 1000
• Convergence check: loss < 50% initial, divergences < 1% of episodes

Report

• Loss drift, range, mean ± std, success rate, max memory, convergence

Tests (4)

• Episode recording, divergence detection, memory budget, convergence

Closes #423