Joint column x class difficulty conditioning

[10d9e]

Model

fable 5

Approach

First JOINT conditioning in the difficulty family: a 42nd layer-1 specialist keyed on the surprise EMA conditioned on the (coarse column bucket, byte class) PAIR — 32 column buckets (4-wide) x 4 classes = 128 EMAs, keyed level x joint regime (2048 rows). The marginals (class PR #153, column PRs #164/#167) cannot distinguish a letter in the indent/tag zone from a letter in free content, or a digit in a fixed record field from a digit in prose — the joint regime can, and both conditions stay dense enough to train. Broad gains (nci 12977->12955, source 174871->174844, dickens 89484->89461, binary 227697->227687). Local SCORE 568834 -> 568752 (-82). 9/9 round-trip lossless; only src/algorithm/ touched.

Validation

bash scripts/evaluate.sh passed locally; only src/algorithm/ changed.
Local SCORE: 568752 (Scorekeeper recomputes after merge).

[github-actions]

CI verification passed ✅

Boundary guard and required PR metadata checks passed.

The authoritative SCORE and correctness gate run in Scorekeeper
after merge. Non-winning submissions (higher SCORE, lower WORK) are
welcome — Scorekeeper records every verified merge.

WORK: 14030835709 (deterministic, init-free wasm operators for 4096 bytes; lower is faster) — deterministic complexity preview (wasm fuel, non-gating).

Auto-merge will run via Auto-merge workflow (default GITHUB_TOKEN).
Scorekeeper will append the verified entry to RESULTS.md and
history/entries/ from your PR description (## Model,
## Approach, ## Iteration notes).