fix: preserve negative zero sign in std.mantissa

[He-Pin]

Motivation

std.mantissa(0 * -1) returned positive 0.0 instead of -0.0. The zero guard used literal 0.0, discarding the IEEE 754 sign bit.

Mathematical basis: std.mantissa implements the mantissa component of frexp (C99 §7.12.6.4). While C says "if value is zero, both parts are zero", IEEE 754 requires sign preservation. All reference implementations (go-jsonnet via Go's math.Frexp, jrsonnet via Rust's bit-preserving (s, 0) return, Python via math.frexp) return (-0.0, 0) for negative zero input.

Modification

Change if (x == 0) 0.0 to if (x == 0) x. Since -0.0 == 0.0 is true in IEEE 754, the guard still catches both zeros but now returns the input unchanged, preserving the sign bit.

std.exponent does not need the same fix — it returns an integer (0L), so there is no sign bit to preserve.

Result

std.mantissa(0 * -1) now returns -0.0, matching all reference implementations. Positive zero behavior is unchanged.

Behavior comparison

Input	go-jsonnet	jrsonnet	Python (math.frexp)	sjsonnet (before)	sjsonnet (after)
std.mantissa(0 * -1)	-0	-0	(-0.0, 0)	0	-0
std.mantissa(0)	0	0	(0.0, 0)	0	0 (unchanged)
std.mantissa(1.5)	0.75	0.75	(0.75, 1)	0.75	0.75 (unchanged)
std.mantissa(-1.5)	-0.75	-0.75	(-0.75, 1)	-0.75	-0.75 (unchanged)

Test plan

• Added mantissa_negative_zero.jsonnet regression test with golden file
• Test uses std.toString to observe sign (assertEqual cannot distinguish -0.0 from 0.0 due to IEEE equality)
• std.toString(std.mantissa(0 * -1)) → "-0" (was "0")
• std.toString(std.mantissa(0)) → "0" (unchanged)
• Full test suite passes (48/48)