orchestrate.py uses global started_at for per-kernel time budget, causing premature termination of later kernels

[yu200512]

Description

In orchestrate.py, cmd_record() computes time_spent_minutes for each kernel using state["started_at"] -- the timestamp when Phase B began (the entire optimization run):

# orchestrate.py line 534-546 (current upstream)
started = state.get("started_at")
if started:
    start_dt = datetime.fromisoformat(started)
    ...
    delta = now_dt - start_dt
    target["time_spent_minutes"] = round(delta.total_seconds() / 60.0)

This means kernel #5's elapsed time includes ALL time spent on kernels #1-#4. With a max_minutes_per_kernel of 120, kernel #5 can inherit 100+ minutes before running its first experiment, hitting the time budget almost immediately.

Reproduce

Run optimization on any model with 5+ kernels:

uv run python auto_optimizer.py --optimize-only ... --iterations 300

Observe:

• Kernel Feature Request: better profiler to identify bottleneck in performance #1: ~20-30 experiments (normal)
• Kernel Integration with 🤗 kernels 👋 #2: ~15 experiments (some time already accumulated)
• Kernel quants support #4-5: 1-2 experiments then "All kernels optimized"

Expected

Per-kernel time budget should start counting from when the kernel begins optimization, not from Phase B start. Kernel #5 should get as much exploration time as kernel #1.

Root Cause

Single global started_at used for all kernels. No per-kernel start timestamp.

Fix

Add kernel_started_at field to each kernel entry, set on transition. Compute time_spent_minutes from kernel_started_at instead of started_at.

---