1.0.2: Robustness

• Reduction-width unification (#754), hardware-write cell separation (#950/#959), Metal's
  device-memory barrier fence (the fence half of #963), matching CPU/GPU half initialization
  (#951), and avoiding tiny repeated CPU pool dispatch (#933).
• Explicit C renderer state (#769), replayable tensor code (#955), wide-index routine logging
  (#953), and partial autotune reports that preserve admitted work (#962/#972).
• Cache, merge-buffer dependency, optimizer, API and precision fixes, see CHANGES.md.
• Test-run launch consolidation and supported queries (#606/#671), verdict provenance
  (#908/#931/#968/#973), shared IR builders (#954), and restored Windows CI setup (#935).
  These improve release verification; they do not close the remaining analysis gaps.
• Code deduplication, mostly in the renderers.