fix: deferred-simp context manager for Ga.build in curvi_linear_latex (#576)

[utiberious]

Companion PR to #590, exploring an alternative code structure for the same SymPy 1.13 fix.

Root cause (same as #590)

SymPy 1.13 PR #26390 added a .replace(Mul, ...) traversal inside TR3/futrig in sympy.simplify.fu. The Ga.build(norm=True) call in curvilinear-coordinate examples invokes Simp.apply ~70 times during metric normalisation, each triggering simplify → trigsimp → futrig → TR3. This scales as O(N·M) for mixed trig+hyperbolic expressions and causes > 10 min hang.

Alternative approach explored here

Two-phase strategy, making the build/output separation explicit in code:

Phase 1 — build: A _no_simp_build() context manager temporarily replaces Simp.modes with an identity simplifier [lambda e: e] around each Ga.build() call. This skips all ~70 redundant Simp.apply traversals during metric normalisation (the intermediate metric components are never printed, so unsimplified storage is fine).

Phase 2 — output: The trigsimp(method='old') profile set in main() is still active when _sympystr calls Simp.apply once per expression at formatting time, avoiding TR3 on the final results.

@contextmanager
def _no_simp_build():
    orig = Simp.modes[:]
    Simp.profile([lambda e: e])  # identity during build
    try:
        yield
    finally:
        Simp.profile(orig)          # restore trigsimp(method='old')

# In each function:
with _no_simp_build():
    (sp3d, er, eth, ephi) = Ga.build(...)

Outcome

	#590 (trigsimp-old profile only)	This PR (deferred build + trigsimp-old)
Build time (prolate spheroidal)	~1.5 s	~0.5 s
Total run time	~5 s	~3.7 s
Output	trigsimp(method='old') canonical form	Identical

Both approaches produce the same output and both fix the CI timeout.

Alternative investigated but does not work

Replacing trigsimp(method='old') with plain simplify() in the output phase (to recover the pre-SymPy-1.13 canonical form) does not help. The unsimplified metric components folded into the gradient/divergence/curl expressions make them large enough that TR3 is just as slow on the final result as on the intermediate metric components. The call hangs after > 10 min on prolate spheroidal grad|A.

Comparison

The structural difference from #590 is that _no_simp_build() makes explicit which Simp.apply calls are skipped (the ~70 build-phase calls) vs which are kept (the ~5 per-output formatting calls). PR #590 achieves the same result via a single outer profile but relies on trigsimp(method='old') being idempotent on already-simplified metric components during build.

Both approaches are correct. This PR exists to let us compare and decide which code structure to land.

Closes #576

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB