Verify real/Double bit-consistency across targets (x87 divergence?)

• Type: feature (verification) — Track A
• Status: DONE — 2026-06-22. Float determinism VERIFIED across x86-64 / i386 / aarch64 / arm32 (no x87 divergence). Durable guard: make test-float-determinism.
• Opened: 2026-06-22
• Closed: 2026-06-22
• Owner: — (Track A / "sis")
• Found by: Track B, building the float math library.

Goal

Confirm the pure-Pascal float math (lib/rtl/math.pas) and the real/Double type produce bit-identical results across all targets. Strict IEEE-754 Double is deterministic on modern CPUs (x86-64 SSE2, AArch64/ARM VFP) and software float emulation, so the results should match exactly — a mismatch is a BUG, not expected divergence. Suspects if any target drifts: legacy i386 x87 80-bit intermediates, or non-strict FMA contraction. (Earlier framing of "floats vary, use tolerance" was wrong — corrected: identical width ⇒ identical result.)

How

Best exact probe: examples/mandelbrot/mandelbrot.pas — its integer escape-count CHECKSUM (3745966 on x86-64, FPC-confirmed) must be identical on every target. Run it under qemu on i386/aarch64/arm32 and diff the checksum. Any difference localises a float-determinism bug. (examples/mathf/mathdemo.pas also runs everywhere; it uses tolerance, so it's a weaker probe — the mandelbrot checksum is the strict one.)

Expected / deliverable

• All targets produce checksum 3745966. If one diverges (likely i386 x87), that's a bug to fix (force strict SSE/64-bit, disable 80-bit intermediates / FMA contraction), not something to paper over with tolerance.
• Record findings; if all match, this closes as "float determinism verified".

Context

Track B verified mathdemo ALL OK on native x86-64 (pinned v35). Cross-target runs are Track A's lane. Float oracles deliberately use tolerance compares, not byte-identical output, precisely because of this open question.

Log

• 2026-06-22 — Filed by Track B from the math-library work.

• 2026-06-22 — VERIFIED, closes. Built examples/mandelbrot/mandelbrot.pas with the current compiler for all four Linux targets and ran the cross ones under QEMU (tools/run_target.sh). Every target produced the integer escape-count checksum=3745966 / ALL OK — identical, so strict IEEE-754 Double is bit-deterministic across x86-64 (SSE2), i386, aarch64 and arm32 (VFP). No i386 x87 80-bit / FMA-contraction drift — the suspect did not materialise. Made it a durable regression guard: new Makefile target make test-float-determinism (builds + cross-runs mandelbrot, asserts the checksum on each target). ESP soft-float (riscv32/xtensa) is covered separately by the softfloat library validation (== x86-64 oracle); not re-run here.