tests: equation()'s 70000-node stack-safety regression test exceeds ctest's 120s TIMEOUT under TSan, failing master's CI

[Yaraslaut]

Split out of #582 — same root cause (equation()'s O(n²) string-building cost, per #582's own measurement table), but a different, narrower fix site and more urgent: this is breaking master's CI right now, independent of whatever #582 decides about a depth cap.

Verification status: reproduced

tests/test_quantity.cpp:736, TEST_CASE("equation() walks a 70000-node provenance chain without overflowing the stack", ...) (added by #581/morph574). Its only assertions (lines 743-747) check output correctness, not performance — 70,000 was chosen solely to safely exceed the old recursive crash threshold (60,000 under clang -O2, per docs/spec/util/quantity_type.md's table). It has no performance budget of its own; it inherits the blanket TIMEOUT 120 every test gets from catch_discover_tests(... PROPERTIES TIMEOUT 120) (tests/CMakeLists.txt:214).

Reproduced on master's own CI, independent of any PR, at commit 3e8f59f2 (the commit that merged #581):

1076/1739 Test #1076: equation() walks a 70000-node provenance chain without
overflowing the stack ...... ***Timeout 120.07 sec

https://github.com/LASTRADA-Software/morph/actions/runs/35431457068/job/105872791321

Also independently reproduced on two unrelated draft PRs (#585, #586) via the pull_request merge-ref — neither PR's own changes touch this code; that's how it was first noticed.

#582 measured 70,000 nodes at 32.3s under ASan+UBSan; TSan's per-access instrumentation is evidently heavier here and pushes the same case past 120s.

Proposed fix (mechanical, does not touch equation())

Give this test (and the sibling destruction test at test_quantity.cpp:723, which is presumably cheaper but shares the same node count and is worth checking) an explicit, longer TIMEOUT via CTest's per-test PROPERTIES, at least under the sanitizer configurations. This is a test-calibration fix, not a design change — it does not require or preempt #582's depth-cap decision, and does not weaken what the test verifies (it still must not crash).

Why this is filed separately from #582

Different fix site: this is tests/CMakeLists.txt's test properties (or the test's own node count); #582 is EquationRenderer::combine in include/morph/util/. Per this repo's own guidance on consolidating issues, two problems sharing a root cause but not a fix site stay separate tickets. This one is also considerably more urgent — master's CI is red on clang-tsan right now.

🤖 Generated with Claude Code

[Yaraslaut]

The stopgap this issue's fix (#590) put in — the [slow] tag and the separate
catch_discover_tests registration at TIMEOUT 600 — is removed again in #603,
which fixes the underlying #582 cost.

Verification status: measured

combine now appends to its left operand instead of copying it into a fresh
string, so rendering a left-leaning chain is linear rather than quadratic in its
depth. Same 70,000-step chain, standalone harness, clang 22.1.8, median of three
runs:

-O0, TSan:        equation=1.245s / 1.260s / 1.381s   (was >120.07s, killed at the deadline)
-O1, ASan+UBSan:  equation=0.112s                     (was 27.720s on master be64026a)

So the test fits the blanket 120 s cap with roughly a 95× margin under the
sanitizer leg that was failing, and tests/CMakeLists.txt is back to a single
catch_discover_tests call.

Two things worth recording, since both are easy to get wrong:

• The test itself had to change, independently of the timeout. Under util: equation() renders an unbounded derivation in full, quadratically — #574's deferred depth cap #582's new
  default step limit equation() stops 100 steps in and never walks deep enough
  to overflow anything, so the test would have gone vacuous — passing against
  the very recursive code it exists to catch. It now passes
  kEquationStepsUnlimited explicitly.
• The TEST_SPEC "~[slow]" filter is removed along with the tag, not left
  behind. A filter excluding a tag no test carries would silently drop any future
  [slow] test from registration.

What would re-open this: the 120 s cap going red again on that test in any
sanitizer leg. The linearity is guarded only by that timeout — no test asserts it
directly — so this issue's failure mode is what would catch a regression in
combine.