feat(bootstrap): host performance model and cycle estimator (Wave 42, R-HS-4)

[gHashTag]

Wave 42 — R-HS-4: Host performance model and cycle estimator

Closes #791

What changed

• New file bootstrap/src/host/perf.rs:

  • EngineConfig — cycles, DMA beats, BRAM utilization, throughput estimation
  • LayerEstimate — per-layer DMA prefetch + compute + DMA drain breakdown
  • PerformanceEstimate — full report across all layers
  • Constants: DATA_WIDTH=54, BRAM_DEPTH=4096, DDR_BEAT=64 bits
  • 19 inline unit tests

• Updated bootstrap/src/host/mod.rs: pub mod perf + re-exports

• Updated bootstrap/src/main.rs:

  • New CLI Commands::HostPerf with --clock-mhz flag (default 66 MHz)
  • run_host_perf() prints single-line estimate

• New file bootstrap/tests/host_perf.rs: 23 integration tests

• Updated docs/NOW.md: wave-42 entry

Test results

77/77 inline host tests pass
25/25 host_driver integration tests pass
21/21 host_irq integration tests pass
20/20 host_engine integration tests pass
23/23 host_perf integration tests pass
= 42 new tests, zero regressions

Output example

$ t27c host-perf --num-layers 3 --neurons 64 --chunks 8 --clock-mhz 66
OK layers=3 neurons=64 chunks=8 total_cycles=3072 total_weight_words=1536 bram_pct=37.5% dma_beats=3072 throughput=21484.4 inf/s @ 66.0 MHz