[enumeratio]

ContinuedFraction(Sqrt(2)) is [1, [2]], ContinuedFraction(ExponentialE, 40) is its 40 terms, and FromContinuedFraction([1, [2]]) is Sqrt(2). Fixes #570

Rules (numerics/continued-fractions.ts, new; library/number-theory.ts):

• An exact quadratic irrational, read from Add, Multiply, Divide, Negate, Subtract, Sqrt, Root(·, 2) and Power(·, 1/2) of rationals and radical literals, expands by the recurrence a = ⌊(P + √D)/Q⌋, P' = aQ − P, Q' = (D − P'²)/Q, with Q | D − P² ensured at the start so the state stays integral. The first repeated state (P, Q) closes the period (Lagrange's theorem), and the first term is always in the leading part: [1, [2]], [1, [1]] for the golden ratio, [1, 7, [1, 2, 1, 8, 13, 8]] for (3 + √7)/5, as Wolfram writes them. With a count n the result is exactly n flat terms. A sum or product of different radicals is not one quadratic field and goes the way of the next rule.
• Another exact real (ExponentialE, Pi, Exp(1/2)) is computed at 2n + 30 digits (_withTransientPrecision), the terms common to both ends of [v − ε, v + ε] are kept less the last (a terminating end writes it two ways), and the same n terms must come out of twice the digits. Not reached within 20000 digits, or the two disagree: unevaluated. n defaults to 20.
• .N() reads the operand as written: it had already made it a float, which has no terms past its precision. The result of .N() is the result of evaluate(), floated the way the wrapper does for a float operand.
• FromContinuedFraction reads a last element that is a list as the period [p₀, …, p_{k−1}]: the matrix of the period gives the tail y as the positive root of C y² + (D − A) y − B = 0, and the leading terms are folded back (x = a₀ + 1/(a₁ + …)). An empty period, a non-positive period term and a non-integer term decline. .N(), and a float term, give a float.
• Caps: MAX_CONTINUED_FRACTION_TERMS (100000) on a period or a count, MAX_CERTIFIED_DIGITS on the precision. Past a cap the call is unevaluated, not truncated.

The signature is (real, integer?) -> list<integer | list<integer>>. A rational, a perfect square and an inexact operand are unchanged (flat list). A negative quadratic irrational uses the floor expansion, as a negative rational does (-√2 is [-2, 1, 1, [2]]); Wolfram's signed terms are not followed.

isqrt in numerics/diophantine.ts is exported for the recurrence. The compile targets have no lowering for ContinuedFraction, so nothing changes there.

Tests: test/compute-engine/continued-fraction-periodic.test.ts (new): ten quadratic irrationals against Wolfram under evaluate() and .N(), n terms, E to 40 and Pi to 30 terms, the cap, and FromContinuedFraction (the same four forms, .N(), declines, and the round trip for every expansion above). Pinned tests changed: wester.test.ts, ContinuedFraction(Sqrt(23)) was twenty terms ending …, 1, 4, now [4, [1, 3, 1, 8]] (and ten terms with a count); the result type of the broadcast ContinuedFraction row in broadcast-collection-valued-results.test.ts, list<list<integer>> to list<list<integer | list<integer>>>.

Overlap: #521, #522, #543, #545 and #547 edit number-theory.ts (other heads; the helpers sit before canEnumerate acceptance, and the CHANGELOG and docs entries will need a rebase); #493 and others edit the same docs tables.

Left out: a float operand's own last term is a truncation of its approximant (ContinuedFraction(2.718281828459045, 40) still ends …, 12, 2); {1, 2, {}} (an empty period) is declined where Wolfram answers 3/2.