Return null from the limit descent's default, not the node (#829, #833)

[Rafael-SOWNet]

Fixes #829 and #833 at the root, and stops six more nodes crashing that #831 does not reach.

The default was the bug

internal virtual Entity? ComputeLimitDivideEtImpera(Variable x, Entity dist, ApproachFrom side)
    => new Limitf(this, x, dist, side);

That is the cycle AGENTS.md names, with this exact expression as its example:

Say "no answer" by returning null, not by handing back an unevaluated node of the expression you were asked about. Limitf(this, ...) looks like the honest answer and is in fact a cycle: the caller evaluates it to compare, evaluating computes the limit, and computing arrives back where it started. That overflows the stack, which kills the process rather than raising anything catchable.

So a node crashed by inheriting the default — by not being mentioned anywhere. Measured on 532d89cf, each of these kills the process:

limit(floor(x), x, 0)     limit(min(x, 1), x, 0)    limit(phi(x), x, 0)
limit(ceil(x), x, 0)      limit(max(x, 1), x, 0)
limit(round(x), x, 0)     limit(gcd(x, 2), x, 0)

The first six arrived with #827 and #828 — mine, and I skipped the ComputeLimitDivideEtImpera override that AddingNode.cs lists at step 3d. phi predates both and is #833, so the default has been a landmine for every node that ever inherited it. That is what #704 was.

The fix is the one AGENTS.md prescribes

Return null. Every caller already expects it: all eleven call sites test with is { } or switch on the result, and the signature was already Entity?. All seven expressions now return an unevaluated limit node — the honest "I could not settle this".

It costs nothing. 5928 tests passed before this change and 5928 after; the extra 45 are the new test below.

The test enumerates rather than lists

LimitTerminatesOnEveryNodeTest reflects over every sealed Entity subtype constructible from a variable — 45 of them — and asserts only that a limit returns, over three destinations and three sides.

That shape is the point. The defect is in what a node inherits by not being mentioned, so a per-node test cannot catch it: the node that crashes is precisely the one nobody wrote a test for. A node added tomorrow is covered the day it is added.

Checked against the unfixed code, where the run aborts rather than fails — a stack overflow cannot be caught, so the signal is the test host dying. Brutal, but detected.

Relationship to #831

This does not replace #831 and does not conflict with it (different files). #831 gives floor and ceil real limit answers and fixes the OverflowException of #830 — it adds capability. This removes the ability to crash. They compose, and the merge order does not matter.

Worth noting that #831 alone would leave round, min, max, gcd and phi still killing the process, which is why this is separate rather than folded into it.

🤖 Generated with Claude Code