Four evaluator arcs: matrix core, ALS-T25 fast-exp, ALS-T9 to_fixed domain, ALS-M15 carrier (abstains 74 to 52)

[O6lvl4]

Stacked on #52. Four requirements-first arcs, each verified byte-identical to
almide 0.59.1 on every burned fixture, with a whole-corpus snapshot diff
proving zero regressions among the previously-evaluated programs.

1 — matrix core (74 → 65): C-161 dims normalize + 2^28 ceiling abort,
C-282 index abort vs reduction identity, C-228/C-229 f32/Q1_0/Q8_0 selectors
and loaders (global-k schedule, per-element bounds, clamps), C-270 full fp16
scale domain + the 0.0 + v accumulate-into-zeros sign ruling, C-223 pow
through vendored musl-libm.

2 — ALS-T25, the canonical fast-exp (65 → 63): new section pinning the
deg-6 unfused Horner algorithm (ties-to-even, 2^k bit shift, ±708 clamp),
softmax row discipline, gelu identity form and associations; C-223 re-points
to it; validation row stamped. The evaluator implements it including the
wasm fmax/fmin NaN-propagating clamp that Rust's f64::max would get wrong.

3 — ALS-T9 to_fixed domain (63 → 59): decimals 0..=4096 through the
exact big-integer expansion, the T6 abort outside, non-finite display forms.

4 — ALS-M15 effect-slot carrier (59 → 52, the class burns to zero):
a callable retagged into an effect (A) -> B position answers
Result[B, String] (pure tail lifted to ok, fallible passthrough). Burning
the class also surfaced and fixed four adjacent gaps, each probed against
0.59.1 first: pipe/?? precedence (parts |> list.first ?? \"none\" reads
(parts |> list.first) ?? \"none\"), in-place list.pop answering
some(last)/none, fs.remove single-entry semantics + fs.is_symlink, and
bytes.to_string answering the ok/err carrier instead of a lossy string.

Ratchets (own commits): 74 → 65 → 63 → 59 → 52. 559/611 programs evaluated.

Author: Claude agent (session als/ref-codec). Verifier: gates only — no
independent human review (QUALIFICATION.md limitation 1).

🤖 Generated with Claude Code

[O6lvl4]

Auto-closed by GitHub when #52's rebase-merge deleted this PR's base branch (ref-codec) — a closed PR's base cannot be retargeted, so this PR's 8 commits ride into main unchanged via #54, which was retargeted to main and rebased (the #52 duplicates dropped as empty patches). Nothing from this PR is lost.

🤖 Posted by the Claude session working PR #52–#54.