spec(port): port specs/port/trios/crates/trios-ternary/src/lib.t27 (Closes #9311)

[dmitrii-f-t27]

Summary

Port gHashTag/trios:crates/trios-ternary/src/lib.rs (at 8b229e9489ee, 175+ lines) as specs/port/trios/crates/trios-ternary/src/lib.t27 — Roadmap Stage 6 (goal #4548), target t27c gen-rust. Closes #9311.

The original is the Ternary BitLinear engine: the {-1, 0, +1} quantization for the bulk compute layers (zero-DSP on the XC7A100T), per the STATIC_ROUTING_TABLE.

Per the issue's rule ("Port the decision, not the plumbing"), all seven listed functions are ported under their names as pure functions (no transports — the growable weight vectors are the callers' plumbing) with 8 tests (28 runtime asserts, 0 vacuous, 100% function coverage):

• Ternary_from_f32 / Ternary_to_f32 — the thresholding decision: strictly above 0.5 → +1, strictly below -0.5 → -1, else 0; and the mapping back (original lines 43-56);
• Ternary_bits_per_param — log2(3), computed without the log2 builtin: a Newton square-root ladder assembles 2^y for any fractional y, and a bisection on 2^x = 3 converges into the 1.584–1.586 window (line 59);
• quantize / dequantize — the scale-then-threshold decision and its division back (lines 76-93);
• compute_scale — max-abs scaling: 1/max_abs, with 1.0 for the all-zero tensor and the empty one (lines 104-115);
• compute_sparsity — the zero share of the quantized tensor (lines 124-127);
• plus the ffn module's decisions: the scale default to compute_scale and the params/5 memory pricing (conservative integer division of the 1.58-bit estimate, lines 137-175).

A note on the second commit

A post-commit dupe_scan found two helper names (sqrt_newton, pow2_frac) already carried by other specs — neither is mandated by the issue — so they are renamed to ternary_sqrt_newton and ternary_pow2_frac, making every introduced name unique across the 2040 specs. Both commits carry their Closes #9311 reference, satisfying the per-commit L1 TRACEABILITY rule.

Verification (all run live from the worktree root, zig 0.16.0, after the rename)

• t27c parse: OK. t27c spec-status: IMPLEMENTED. t27c typecheck: 0 errors, 0 warnings.
• t27c coverage: 14/14 tested (100%) — no transports in this port.
• t27c test-report: 8 tests / 8 pass / 0 FAIL / 0 BLOCKED, 0 vacuous passes — the threshold boundaries (0.5 and -0.5 stay 0), the log2(3) window, the sqrt ladder windows (√2, ⁴√2), the quantize/dequantize round trip, both scale fallbacks, the sparsity shares, and the params/5 pricing with its conservative truncation.
• t27c gen | grep -c 'not yet implemented': 0; gen | wc -l: 169 (> 12).
• grep -cE '^\s*(pub )?fn (Ternary_from_f32|Ternary_to_f32|Ternary_bits_per_param|quantize|dequantize|compute_scale|compute_sparsity)\(': 7. grep -cE '^\s*test\s+(["|[A-Za-z_])': 8 (≥ 7).
• Boundary: git status --short — exactly this one file. The repo's lefthook pre-commit passed on both commits.
• dupe_scan: all introduced names unique across 2040 specs (after the rename); the bare body scan adds no duplicate group (752 bodies / 221 groups on base, identical with this spec present).

Dmitrii Fedorov (@dmitrii-f-t27)

[dmitrii-f-t27]

@gHashTag Please merge this PR into master via the golden chain — rights on gHashTag/t27 are read-only for me (403), so the merge itself needs your action.

What this is: the #9311 port — crates/trios-ternary/src/lib.rs (gHashTag/trios, at 8b229e9489ee, 175+ lines) re-authored as specs/port/trios/crates/trios-ternary/src/lib.t27. One file, head 6c3324da7, Closes #9311. The second commit renames two helper names (sqrt_newton, pow2_frac) a post-commit dupe_scan found carried by other specs — neither is mandated by the issue; both commits carry their Closes #9311 reference, so the per-commit L1 TRACEABILITY rule holds — Check L1 is green.

What is verified:

• t27c parse OK, spec-status IMPLEMENTED, typecheck 0 errors / 0 warnings, coverage 14/14 (100%) — no transports in this port, t27c test-report 8 tests / 8 pass / 0 FAIL / 0 BLOCKED, 0 vacuous — the {-1, 0, +1} thresholding with its boundaries (0.5 and -0.5 stay 0), the value mapping back, bits_per_param = log2(3) computed without the log2 builtin (a Newton square-root ladder assembles 2^y, a bisection on 2^x = 3 converges into the 1.584–1.586 window), the quantize scale-then-threshold decision and its dequantize division, max-abs scaling with both 1.0 fallbacks, the sparsity zero-share, and the ffn params/5 pricing with its conservative truncation.
• gen → 169 lines, 0 not yet implemented; the 7 required names count exactly 7; 8 test blocks; boundary exactly the one file.
• dupe_scan: all introduced names unique across 2040 specs (after the rename); the bare body scan adds no duplicate group (752/221 unchanged).
• The t27b-native-run artifact of this head (commit 6c3324d, run 39018042696): lib.t27 row — reference pass, t27b pass, tests 8, invariants 0, asserts 34, blockers [], detail empty, reference_disagree: [].

CI (head 6c3324d): required trio green — validate, check-linked-issue, parse-ratchet — Check L1 TRACEABILITY green and Corpus ratchet green; 26 checks pass overall, incl. emit-bitexact, t27b-native-linux, t27b-native-macos.

The 3 reds are the base-red trio, not this change: cli-tri (the #3324 include_str!), duplicate-bodies (only the known base names), t27b-native-ratchet (the same verdict red the Queen's port PRs carry; this spec's own row is reference pass + t27b pass — the registration debt is the UNLISTED class).

After the merge I will: verify the spec file live on the merge SHA, confirm #9311 auto-closed, and file the 14-step report.

Dmitrii Fedorov (@dmitrii-f-t27)