feat(spec): BitNet×GF-T inference neuron + signed GF-T dot4 (Refs #1764)

[gHashTag]

Refs #1764

Two spec-first increments fusing the project's two threads (BitNet ternary + GF-T format), both bit-exact to the ideal oracle gft16_ref.py and synthesizable. No compiler change (uses on_comb).

BitNet×GF-T neuron — specs/ternary/gft_bitnet_neuron.t27

A neuron with ternary weights {-1,0,+1} and real-valued GF-T16 activations — the actual BitNet inference pattern with wide dynamic range. Each trit weight selects +a / 0 / -a of its GF-T activation; the four signed contributions are summed in signed GF-T (RNE, zero-aware).

• Bit-exact to the oracle over 3000 vectors (tree-consistent adds).
• yosys synth_xilinx → 9763 LUT + 1320 CARRY4 (synthesizes to Artix-7).

Signed GF-T dot4 — specs/ternary/gft_signed_dot4.t27

A signed GF-T16 4-term MAC (real-valued matmul tile with negatives + cancellation), bit-exact to the oracle balanced tree over 3000 vectors.

Synthesis finding

A bounded while loop inside a Verilog function is not yosys-synthesizable ("Function can only be called with constant arguments"). The signed subtract's left-normalization was rewritten as a flat unrolled sequence (12 conditional shifts, ≥ the ~9 max) — functionally identical (re-verified 3000 vectors), now synthesizable. Bitstream/place-and-route (nextpnr) remains owner-gated (not installed locally).

1535 unit tests pass.

🤖 Generated with Claude Code

[github-actions]

📓 NotebookLM Notebook linked to this PR

• Notebook ID: b83263109fb055dc
• URL: https://notebooklm.google.com/notebook/b83263109fb055dc

This notebook contains session context, decisions, and artifacts for this work.

[github-actions]

PR Dashboard

Generated at: 2026-08-06 14:24:07 UTC

• feat(spec): BitNet×GF-T inference neuron + signed GF-T dot4 (Refs #1764) #1795 feat(spec): BitNet×GF-T inference neuron + signed GF-T dot4 (Refs #1764) | Created: 2026-08-06 | Updated: 2026-08-06 | 11 failing / 15 total
• fix(codegen): restore balanced translate_off/on guards and reseal all specs #1793 fix(codegen): restore balanced translate_off/on guards and reseal all specs | Created: 2026-08-06 | Updated: 2026-08-06 | 0 failing / 3 total
• fix(codegen): faithful Rust/C emit — don't run optimizer on source backends (#1455) #1456 fix(codegen): faithful Rust/C emit — don't run optimizer on source backends (#1455) | Created: 2026-07-10 | Updated: 2026-07-10 | 3 failing / 30 total

Summary

Status	Count
Total Open PRs	3
PRs with Failing Checks	1
PRs with All Checks Green	2
READY	1
FAILING	1
PENDING	0

Seal Status

• ⚠️ STALE -- sha256(compiler.rs)=de57378ec0e2 != manifest seal=87e5cbd3ad94.
  The committed NMSE numbers were certified against an older compiler.rs.
  Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).