Float math library (Single+Double) + numerical float demo

• Type: feature — Track B
• Status: done — green on pinned v35 (commit a25136d)
• Opened: 2026-06-22
• Owner: —

Goal

A comprehensive floating-point math library (the foundation a float demo needs), plus a float demo/oracle. Pure-Pascal fallback, native x86-64, no asm yet (other-CPU goniometric asm comes later). Single + Double overloads only — no Extended (it is aliased to Double; see feature-extended-type-support).

Delivered

• lib/rtl/math.pas expanded: Tan, ArcSin/ArcCos/ArcTan2, Sinh/Cosh/Tanh + inverses, Cot/Sec/Csc, Log10/Log2/LogN, Hypot, IntPower, Floor/Ceil/FMod, Sign, float Min/Max, DegToRad/RadToDeg — on top of the existing Pi/Sqrt/Exp/Ln/Sin/ Cos/ArcTan/Power core. Single overloads (widen→Double→narrow) for the transcendental/utility set. Reuses builtins Trunc/Round/Frac/Int.
• examples/mathf/mathdemo.pas — tolerance oracle + numerical showcase: every function vs textbook constants (1e-9); Single overloads (1e-5); Single→Double conversion; mixed single/double promotion (resolves Double); the real type; Machin π / e-series / Simpson ∫sin / Newton √2. Ends ALL OK; wired into make lib-test + make demos.

Foundations verified (compiler, on v34/v35)

• Single→Double conversion and narrowing assignment: OK.
• Overload resolution Single vs Double: OK.
• Mixed-type promotion single*double → double: OK.
• real behaves as Double on x86-64. Cross-target consistency: separate ticket feature-real-cross-target-consistency (Track A, qemu runs).

Gaps found (filed)

• bug-untyped-float-const: const X = 1.5; rejected (typed const works); also Single(expr) value cast not parsed. Worked around with clean idiomatic Pascal.

Determinism

Floats are not guaranteed byte-identical across targets, so the oracle uses tolerance compares + ALL OK, never raw float output.

Log

• 2026-06-22 — Implemented + green on pinned v35 (a25136d). Math library is the foundation; float demo built on it. Done.