[Rafael-SOWNet]

Fixes #2166.

The defect: a sum simplifies its left operand first, and that operand's left operand, and so on. So a chain of n sums nested to the left was simplified by a recursion n deep. A sum built one term at a time has that shape: CollectLikeTerms builds one, and Simplify's search hands the collected chain on to later candidates.

The integral of (q + p x)/(c + b x + a x^2)^3, with its symbols pinned, reached about 900 levels:

• On most Windows runs since 2026-09-15 it overflowed C# Test's Debug build on a 1 MB stack. The Windows test leg runs the whole suite, in Release, and fails when it does not #2165 moved that leg to Release, and the traces are listed there.
• In Release it needed more than 384 KB, and macOS gives a secondary thread 512 KB.

The change: Sumf and Minusf first simplify every 32nd node of a left spine that is at least 64 deep, deepest first. Every node's simplification is cached, so each of those recursions stops at the checkpoint below it. The same nodes are simplified, each once, in a different order. An ordinary sum pays only for a walk of a few nodes.

Measured on a thread of a given stack size:

case	master	this PR
cubic row, Release, 384 KB	overflow, 909 repeats	passes
cubic row, Release, 256 KB	overflow	passes
cubic row, Debug, 1 MB	overflow, 654 repeats	passes
(a_1+..+a_30)(b_1+..+b_30) expanded, Release, 256 KB	overflows at 400 terms	900 and 1,600 pass

The next thing to overflow on the cubic row, at 192 KB in Release, is CostModel.ChildrenCost, 872 deep. That is a different recursion, and it has 2.7× headroom on macOS.

Test: LongSumStackDepthTest expands the 900-term product on a 256 KB thread.

Suite and performance:

• The unit suite passes: 17,128 passed, 0 failed.
• Against master, on one machine, alternating the two builds:

	master	this PR
InnerSimplified on five ordinary sums, bytes per round	23,356	23,356
the same, ms per 200,000 rounds (mean of 3)	2,525	2,557
Simplify on ten expressions, bytes per round	16,745,849	16,745,849

The time difference is +1.3%, inside the 5% spread between runs of a single build.

No answer changes, so there is no BREAKING-CHANGES entry.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.41%. Comparing base (90ac975) to head (c25df29).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2167   +/-   ##
=======================================
  Coverage   86.41%   86.41%           
=======================================
  Files         295      295           
  Lines       64790    64806   +16     
  Branches    19053    19059    +6     
=======================================
+ Hits        55987    56005   +18     
+ Misses       4773     4770    -3     
- Partials     4030     4031    +1     

Flag	Coverage Δ
cpp	37.77% <ø> (ø)
csharp	86.99% <100.00%> (+<0.01%)	⬆️
fsharp	32.97% <ø> (ø)
interactive	40.72% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
...nuous/Evaluation.Continuous.Arithmetics.Classes.cs	87.48% <100.00%> (+0.32%)	⬆️

... and 12 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.