[enumeratio]

Fix #434.

The JavaScript target now follows the compile mode for these heads, as Sqrt, Ln, Arcsin and Power already do:

• In auto and complex mode: where the facts about the operands don't prove the value real, the head takes the complex lane. It calls the interpreter's complex kernel through _SYS, the way Gamma uses _SYS.cgamma.
• In strict mode: the real kernel stays, and a complex value gives NaN.
• A complex operand takes the complex kernel in every mode.

The fact that decides "real, complex or unknown" is in one place, complexContinuation (compilation/complex-continuation.ts). Both isComplexValued and the emitters read it.

head	real kernel kept when	complex kernel
HurwitzZeta(s, a)	a ≥ 0 or integer s	_SYS.churwitzZeta
LerchPhi(z, s, a)	z ≤ 1 (or integer s ≤ 0), and a ≥ 0 or integer s	_SYS.clerchPhi
PolyLog(s, z)	z ≤ 1 or integer s ≤ 0	_SYS.cpolyLog
EllipticK(m), EllipticE(m)	m ≤ 1	_SYS.cellipticK, _SYS.cellipticE
Hypergeometric2F1	z ≤ 1, or the series terminates	_SYS.chypergeometric2F1
LambertW(x, k)	x ≥ −1/e on branch 0; [−1/e, 0) on branch −1	_SYS.clambertW

GPU: HurwitzZeta, LerchPhi and PolyLog decline to compile when the value is provably complex (e.g. LerchPhi(3, 2.5, 1)). When it is only unknown, the real helper stays, so a shader over plain symbols still compiles. The other heads have no GPU lowering.

Arcsec and Arccsc: over a real operand of unknown magnitude they are now typed complex | infinity | nan, where before they were typed number. They then take the complex lane, as Arcosh does. Compiled Arcsec(0.5) was NaN and is now 1.3170i.

Tests:

• The NaN pins in zeta-values, polylog-order, lerch-phi-values and item-374-hurwitz-negative-order now expect the interpreter's value.
• A new table-driven compile-complex-continuation.test.ts checks each head:
  • compiled in auto and complex mode against .N();
  • strict gives NaN;
  • the real region is unchanged in every mode;
  • a complex operand takes the complex kernel;
  • the GPU targets decline a provably complex value.

Left out: incomplete EllipticE and StieltjesGamma keep their real-only lowerings.

[arnog]

Merged by hand onto main as 403a216, with the CHANGELOG/ROADMAP conflict resolved. Two additions: a positive integer order on the cut z > 1 is also provably complex, so PolyLog(2, 3) and LerchPhi(3, 2, 1) decline on the shader targets too; and a user definition with its own compile handler keeps its own lane. Thank you.