Wire %fixp (fixed-point) arrays into Lagoon

[sigilante]

Phase B of the fixed-point buildout — integrates /lib/fixed into Lagoon as
the %fixp array kind, element-wise plus exact reductions, mirroring the
%int2/%unum integrations.

Independent of PR #43 (Phase A scalar API): this uses only /lib/fixed's
pre-existing mul/div/scale/to-s/ng arms.

Representation

A fixed-point element occupies 2^bloq bits, with a + b + 1 = 2^bloq
(q3.4 @ bloq3/8b, q7.8 @ bloq4/16b, q15.16 @ bloq5/32b) — the same power-of-2
tiling every other kind uses. The precision [a b] rides in meta.tail
(its declared "future data" slot), read via ;;([a=@ b=@] tail).

What's wired

• sur/lagoon: %fixp enabled in the kind union.
• fun-scalar: add/sub/mod and all comparisons need only the
  element width and reuse /lib/twoc directly (identical to %int2 on the
  stored integer). mul/div rescale by 2^b through /lib/fixed; mul
  re-quantizes back to the array precision rather than widening (array
  elements are uniform — truncating, documented). Comparisons return the
  fixed value 1.0 (= 2^b) / 0.0.
• trans-scalar %abs: two's-complement abs at the element width.
• get-term (%ux, no fixed-point printer), eye/ones constant
  1.0 = 2^b from meta.tail, scale raw-pack, change guarded
  with a ~| message (conversion not yet wired).
• Reductions: dot/mmul use a new +fixp-fdp that accumulates the
  integer products Σ sx·sy exactly, then rescales by 2^b once — exact,
  no per-product truncation. cumsum needs no special path (fixed add is
  already exact modular integer add).

Tests — /tests/lib/lagoon-fixp (7, all pass on ~hex)

q3.4 coverage: arithmetic, comparisons, abs, the ones builder, cumsum,
and exact dot/mmul. The %unum (14) and %int2 (5) regression suites
remain green.

Not in scope

change/convert for %fixp (use /lib/fixed's to-rs/from-rs); jets.

🤖 Generated with Claude Code