gft encoder: pin round-half-to-even on exact ties (Refs #2161)

[gHashTag]

if t > hf: mant += 1 in _magadd survived the boundary operator. It is the round-half-to-even decision:

if t > hf: mant += 1                  # strictly above half
elif t == hf and (s & 1): mant += 1   # exactly half -> to even

With >= the pair becomes round-half-up and the parity test on the next line is dead code. Measured, not predicted:

_magadd(25600, 20480)   clean 25600    mutated 25601
_magadd(20992, 20481)   clean 21248    mutated 21249
_magadd(21504, 20482)   clean 21632    mutated 21633

256 such witnesses exist, and the self-tests notice none of them.

Why an outcome test could not see it

The self-tests train a network and assert on accuracy. An optimiser absorbs a last-bit error in every addition without changing whether XOR reaches 4/4 — planted, the mutant passes every training assertion in the file.

A test that checks an outcome cannot see a defect the outcome tolerates. Accuracy, throughput, "it still converges" — each is a real property, and each is a filter that removes exactly the errors an adaptive process routes around. Arithmetic needs assertions on arithmetic.

Three assertions now pin the decision, one per branch. Negative controls: weakening > to >= fails the even-tie assertion; disabling the parity test fails the odd-tie one. Each control hits the assertion aimed at it, which is what makes them three tests rather than one repeated.

Column: 6/31 → 7/31.

And a correction

This tick was planned around microcode:203 (if d < 0 in _magsub), carried from yesterday's classification table. It is not a survivor — with <= the XOR self-test reports 1/4 and the mutant dies loudly.

I had found it by grep, classified it as a real boundary, then silently promoted "real boundary" to "surviving boundary". Those are different claims: one is about the code, the other about what the tests reach.

§70 is mine — a finding recorded as a line number expires. This is worse: it never was a finding, and the line number gave it enough shape to look like one. The tick recovered only because the first thing I did was reproduce the survivor rather than fix it.

§112 and §113.

Refs #2161

[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: 2026-08-24 00:58:49 UTC

• gft encoder: pin round-half-to-even on exact ties (Refs #2161) #2640 gft encoder: pin round-half-to-even on exact ties (Refs #2161) | Created: 2026-08-24 | Updated: 2026-08-24 | 18 failing / 25 total
• Unbreak FPGA CI and correct two unbacked hardware claims #2081 Unbreak FPGA CI and correct two unbacked hardware claims | Created: 2026-08-11 | Updated: 2026-08-20 | 5 failing / 27 total

Summary

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

Seal Status

• ⚠️ STALE -- sha256(compiler.rs)=e6333575ec10 != 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).