Numeric silent-wrong mining: differential corpus vs Julia/R/NumPy (before #1644 numbers phase)

[InauguralPhysicist]

Mine the numeric semantics that Julia, R and NumPy settled deliberately long ago, to find EigenScript's silent wrongs before the numbers phase of #1644 decides them. This is a differential corpus, not a guest-language port: ports of near-home languages yield few gaps, but these languages solved exactly the numeric edge cases where our silent wrongs keep turning up (#1649 sum skips non-numbers, the fail-soft audits, the AOT's aot_num_or(..., 0.0) sites).

Owner request (2026-10-07): log it so it doesn't get forgotten. Run it after the current bool/EigenLua work, ideally as a cloud run. The oracles (Julia, R, NumPy) are installed only in the sandbox; no Python enters any repo.

Edge-case axes, per numeric builtin

• NaN in reductions and ordering: min/max/sort/mean/sum (NumPy max vs nanmax; R's NA vs NaN).
• Empty reductions: sum([]), mean([]), max([]), min([]).
• inf, -inf and -0.0: arithmetic, comparison, printing, use as a key.
• Integer overflow and precision past 2^53 (Julia wraps, with checked types; R gives NA plus a warning).
• Float printing: shortest round-trip, -0.0, rounding mode (half-even).
• Mixed types: number + bool/string/null in every numeric builtin. It must raise, never skip or coerce.
• Missing-value propagation (R's NA rules as a design reference).

Done when

• A corpus covers every numeric builtin from eigenscript --api --json against each axis above, with each case's output recorded from EigenScript, Julia, R and NumPy.
• Every disagreement is triaged as either (a) a silent-wrong issue filed with a reproducer, or (b) a deliberate EigenScript choice written into docs/SPEC.md with its reason.
• The results are linked from Value-model release: booleans, then integers, strings and map keys in ONE breaking release #1644 before its numbers phase gets an owner decision.