Integration and limits with the downcasting off compute as with it on, on exact numbers

[Rafael-SOWNet]

Closes #1490.

With DowncastingEnabled off, integration and limits gave NaN, or left unevaluated, what the default setting answers. There were two causes:

• The zero test. TreeAnalyzer.IsZero compares against the literal 0, and with the downcasting off a zero is Complex(0, 0), which isn't equal to it. So nothing was zero, and 1/(1 + c^2 x^2) kept its a = 0 case, ln(1)/0.
• The parse. A whole number parsed with the downcasting off is a decimal Real, so x^2 isn't a whole power to the integrator.

Integration and limits now compute as with the downcasting on, whatever the caller's setting, on the input's decimals read as the exact rationals they are. The answer no longer depends on the setting. The cost, as proposed on the issue, is that decimals come back as rationals: the integral of 0.1 x is x^2/20.

With the downcasting on, nothing changes and nothing is added to the path: the setting is read, and that's all. With it off, the scope is opened once, at the outermost call, since every call inside it finds the downcasting on.

Fixing the zero test alone would have fixed the NaN, but not the rest. The rules still test numbers by type and against literals, and with only the zero test fixed, (d - c^2 d x^2)^(3/2) (a + b arcsin(c x)) gave up after 290 s instead of 7.

with the downcasting off	master	this PR
(a + b arcsin(c x))/sqrt(d - c^2 d x^2), parsed with it on	NaN + C	the antiderivative
(d - c^2 d x^2)^(3/2) (a + b arcsin(c x)), parsed with it on	unevaluated, after 6.2 s	the antiderivative, in 0.13 s
1/sqrt(2 - 2 x^2), parsed with it off	unevaluated	-arcsin(-x) sqrt(2)/2, up to form
x^2 sin(x), parsed with it off	unevaluated	the antiderivative
limit(sin(c x)/x, x, 0)	NaN	c
limit((1 - cos(c x))/x^2, x, 0)	NaN	c^2/2
0.1 x, parsed with it off	0.1 * x^2 / 2 + C	1/10 * x^2 / 2 + C

Not in this PR

The parse cache is keyed by ExplicitParsingOnly but not by DowncastingEnabled. A string parsed under one setting therefore comes back under the other with the first parse's numbers. That's #1513.

Measured

• The unit suite passes on net10.0: 12883 passed, 14 skipped, none failed, of 12897.
• The new tests bind: 8 of the 9 in DowncastingOffCalculusTest fail with the library change reverted. The ninth is 1/sqrt(2 - 2 x^2) parsed with the setting on, which master already answers, and it is there to keep that so.
• The allocation gate passes on all 19 gated benchmarks. None moves.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura