CI receipt: whole Living Weights neighbour subjects (#714)

[cwschroeder]

This draft runs the upstream-format and no-GPU workflows for the exact upstream candidate commit. It is a CI receipt for upstream draft #714. Both fork checks passed on 9b52ed0; upstream Actions require maintainer approval.

Living Weights reuses a neuron's single-word key when it generates questions about other entities. With Project as the key for Project Aster, the model is asked for alternatives to Project, producing questions such as “What is Task Aster's incident code?” The router also applies the first retained question's subject to every wording, and both replacement paths can match inside longer names.

This carries a separate replacement span for each retained question. It extends the existing key over adjacent, already identified name words, so Project Aster and Quillmoor Labs are replaced whole. Both generation paths use the same complete-word matching and preserve possessive suffixes. The router uses each wording's own span and caches alternatives per exact span.

Neuron keys, token positions, positive-question selection, numerical kernels, recall judging and rollback rules stay as before. A wrongly selected key such as Quality still stays Quality; this patch does not correct that earlier selection. Model-generated alternatives can still be unsuitable. More distinct spans can require more model queries; the existing 96 limit bounds accepted router examples, not all queries.

Validation

• Zig 0.17.0, complete zig fmt --check build.zig zig passes.
• New CPU target zig build test-slide-neighbours passes 23/23 tests. It also runs inside the ordinary zig build test on Linux. Cases cover multiword names, different subjects, null spans, punctuation, whitespace, case, ASCII/Unicode possessives and Aster versus Asterion.
• zig build native test -Dcpu=apple_m1 -j2 --summary all: 143/143 steps succeeded. Initial implementation db69e7a ran 1,296 tests with two skips; head 9b52ed0 adds diagnostic logging and reran 260 affected tests with the remaining dependencies cached.
• Upstream base is e7c523ebfa16b8094cf9e35e3773a7d25befcc5b; native binary SHA256 is 63cb13b8824f0c69b286d56ac1907cf38fddf8729e990b8b66900ceea9ba1a79.

The frozen single-fact A/B run on Apple M5 Max / 128 GiB completed (48 replies per arm, 96 total). Both arms used fresh copies of the same untaught Nemotron checkpoint, identical initial replies, one fictional document fact, German/English questions, thinking off/on and a real restart.

After restart	Stock upstream	This candidate
Strict recall	1/4	1/4
Neighbour correctness (baseline 8/8)	4/8	4/8
Controls	4/4	4/4
Lesson seconds including save	551.74	619.75

Both lessons were accepted. The candidate records 18 retained questions with neuron key Project and replacement span Project Aster, and generates 12 same-code questions about other projects; stock retained only a Task Aster alternative. None of the exact external questions appears in the learner logs. Visible answers and finish reasons survive restart, original weights/config stayed unchanged and all owned test children exited. This fixes the replacement mechanics; this run provides no improvement in external recall or aggregate neighbour correctness. Timings are observed experiment costs. Reviewed responses, receipts and independent audit record these failures and all build/corpus pins. Native/Hugging Face tokenizer parity also confirms that the complete Aster token sequence is still a shared prefix of Asterion, so a later learning anchor needs separate evaluation.

A separate existing-policy arm used the identical candidate binary and frozen inputs with only TF_SLIDE_POCKET_KEY=reply. Actual allocation was 10 Reply / 0 Subject / 0 Value neurons. After five unsuccessful rounds the learner rejected and removed the lesson. Its 48 external replies, including a genuine restart, return to the original full baseline: recall 0/4, neighbours 8/8, controls 4/4, all finished normally. The neighbour preservation is rollback, not an accepted learning improvement. Reviewed C responses, receipts and independent audit record this negative result. No further native source changes or production deployment were made.

Current-head fork CI passed formatting and no-GPU tests on 9b52ed0; upstream Actions remain action_required awaiting maintainer approval.
Related: #646. This has no dependency on the client-tools proposal ashhart#664 or the supplied-question proposal ashhart#687.