sim-rs: Add timestamp quantization and TX batching to sequential engine

[sandtreader]

Summary

The sequential DES engine was processing events almost entirely single-threaded during TX-only periods (near 0% CPU on a 32-core machine), only bursting to ~6 cores after EB production. Two root causes:

1. No timestamp quantization — the sequential engine never applied timestamp-resolution-ms. Every TX arrival and network delivery had a unique timestamp, so the BSP batching processed ~360M events one at a time, never reaching parallel-threshold for rayon.

2. One TX per event — TxGeneratorCore::generate() produces one TX per call, each scheduled at a unique time. Even with quantization, TX inter-arrival times may span multiple quanta.

Changes

Timestamp quantization in the sequential engine (sim-core/src/sim/sequential.rs):

• All event timestamps pushed to the priority queue are now quantized via Timestamp::with_resolution() — network deliveries, CPU task completions, timed events, and cross-shard messages.
• This causes events with nearby timestamps to land in the same time bucket, creating larger batches that trigger rayon parallelism.

TX batch generation (new tx-batch-window-ms config option):

• When set, the TxGeneration handler generates all TXs whose next scheduled time falls within the configured window, instead of one per event.
• Independent of timestamp-resolution-ms — works by comparing against an absolute window rather than quantized timestamps.

turbo.yaml tuning:

• timestamp-resolution-ms: 1.0 (was effectively unused at 0.000001)
• parallel-threshold: 4 (was 10)
• tx-batch-window-ms: 10.0 (new)

Performance impact

Tested on a 32-core machine with a 1500-slot CIP simulation:

• Before: ~90 min wall clock, 0.2% CPU during TX-only periods
• After: ~12 min wall clock, 70% sustained CPU utilization
• ~7.5x speedup

Backward compatibility

All three optimizations are opt-in via turbo.yaml. Without it, defaults match previous behavior exactly:

• timestamp-resolution-ms: 0.000001 (no effective quantization)
• parallel-threshold: 10 (unchanged)
• tx-batch-window-ms: null (disabled)
• Default engine is actor, not sequential

Test plan

• All 41 existing tests pass
• Clippy clean
• Verified default parameters (without turbo.yaml) match previous behavior
• Benchmarked with turbo.yaml on 1500-slot CIP simulation

🤖 Generated with Claude Code