Add floor and ceil (#809)

[Rafael-SOWNet]

Part of #809 — the two least contested of the six functions it asks for. round, min, max, gcd and lcm are untouched and still refused by name.

Every convention was measured, not chosen

AGENTS.md asks for this when a convention has to be settled, so the values come from SymPy 1.14 rather than from reasoning. Three of them are not what the issue assumed:

Rounding direction	toward the infinities, not toward zero — floor(-3/2) is -2, ceil(-3/2) is -1
Complex argument	componentwise: floor(3/2 + 5/2i) is 1 + 2i. SymPy and Mathematica agree, and it is the only reading under which floor of a real keeps its meaning when the imaginary part happens to be zero
Derivative	0 provided not x in ZZ. The issue guessed "zero almost everywhere"; SymPy declines to answer at all. Stating the condition says more than either, and it is exactly what Signumf and Absf already do — this library has Providedf and SymPy does not

ceiling( is accepted too, since that is SymPy's spelling and an expression copied from there should parse. Stringize prints the short ceil(, which is what the round-trip test pins.

Inverting them

floor(f(x)) = v is solvable only where v is an integer, and then f(x) is anywhere in the half-open interval [v, v + 1) — so the inverse is a parameter over that interval, not a point. That is the same device Signumf and Absf use for their own many-to-one inverses.

floor(x) - 3 = 0    ->    { 3 + t_1 provided t_1 in RR and t_1 >= 0 and t_1 < 1 }

#733 recorded that same equation answering { 3 / floor } — a root of nothing. It is now answered correctly, and there is a test asserting it.

The grammar

Regenerated with JDK 25 and the committed ANTLR 4.13.1 jar, followed by the post-processing step. The unmodified grammar was regenerated first and reproduced the checked-in parser byte for byte, so the diff here is the new rules and not a toolchain difference — which is the check Docs/Contributing/ImproveParser.md asks for and the only reason this was safe to touch.

Three lines moved out of the not-implemented list and into the implemented one.

Tests

5872 passing, 0 failed, 14 skipped. F# 130 passing.

FloorCeilTest covers values against SymPy's answers (including every negative case), the complex componentwise rule, idempotence, symbolic arguments left alone, the derivative, LaTeX, the printed form parsing back, substitution, and the solve.

Seven existing tests failed and were changed, and it is worth being explicit about which kind that is: they asserted floor/ceil/ceiling are refused, and they no longer are. That is the answer getting better, not a test pinning a fudge. The refusal cases move to names that are still missing, and one gains a counterpart asserting the original report's equation is now answered. A new test covers the neighbouring case — an implemented name with the wrong argument count now says the count is wrong rather than that the name is unknown.

Compatibility

BREAKING-CHANGES.md records it: these three spellings used to raise UnrecognizedFunctionParseException and now return expressions. If you were catching that around these names, it is dead code. floor without a bracket is still an ordinary variable, and the other refused names are unchanged.

Not in this change

round needs the tie-breaking convention writing down — measured as half-to-even in SymPy, Python and IEEE 754, but not what Math.Round(decimal) does, so it deserves its own note. min/max want the node-versus-sugar decision, on which SymPy keeping them as nodes is evidence. gcd turned out to be the polynomial gcd rather than an integer one, and this library already has PolynomialGcd — so it should be looked at alongside that rather than as new number theory. All three are recorded on #809 with the measurements.

🤖 Generated with Claude Code