Improve telemetry percentile accuracy with reservoir sampling

[lambda-alpha-labs]

The telemetry module uses Exponential Moving Average (EMA) for P50/P95/P99 percentile estimation. While lightweight, EMA is order-dependent: the same data processed in different order produces different percentile estimates.

Proposed fix

Implement reservoir sampling: maintain a sorted buffer of the last 100-200 latency measurements per node. This ensures order-invariant percentile calculation while still being memory-efficient.

Current code

src/telemetry.rs in update_percentiles:

meta.p50_ms = meta.p50_ms * 0.9 + latency_ms * 0.1;

Acceptance

• Percentile values are invariant to trace file ordering
• Memory overhead stays under ~1KB per active node
• No regression in existing telemetry tests