Validation

[Preston-Walker]

Phase 1: widen cell-count decrease threshold from 3% to 50% in the structural
checker to avoid false negatives when Vivado optimization passes legitimately
reduce cell count.

Phase 2 simulation overhaul:

• Replace the dead "paired" testbench mode with a trace-replay architecture:
  golden design writes stimulus+outputs to a .hex trace; revised design replays
  the same stimulus and outputs are compared. Removes the paired testbench
  template entirely.
• Add a content-addressed xelab snapshot cache (keyed on SHA-256 of the golden
  DCP + validator script) so repeated validation runs skip slow elaboration.
  Cache directory excluded via .gitignore.
• Replace hardcoded LFSR seed (0xDEADBEEF) with a per-run random seed drawn
  from os.urandom(4). Seed is passed to xsim at runtime via +LFSR_SEED so
  xelab snapshots remain seed-independent and the cache continues to work.
  Seed is encoded in the cached trace filename (seed_XXXXXXXX_vectors_N.hex)
  so a cache hit naturally reuses the same seed.
• Add --seed CLI flag for reproducible/deterministic runs during development.
• LFSR_SEED testplusarg is now only passed for golden trace generation, not
  revised replay (which has no LFSR and ignored it).