Fix lambdify cache-miss on evaluate() (closes #194) + replace flaky timing tests

[lmoresi]

Summary

Two coupled changes in the lambdify-evaluate caching area:

1. Fixes uw.function.evaluate() never hits the lambdify cache (fresh sympy.Dummy per call → +1 cache entry every call) #194 — uw.function.evaluate() never hit the lambdify cache. _expr_hash used sympy.srepr(expr), which embeds the volatile global dummy_index of any sympy.Dummy. The evaluate() coordinate-substitution path mints a fresh Dummy per call, so an otherwise-identical expression hashed differently every call → _lambdify_cache grew one entry per call (1,2,3,4,…) and sympy.lambdify recompiled every time on a hot path.

   Fix: canonicalise Dummy → Symbol(d.name) in _expr_hash before srepr. This changes only the cache key; the real sympy.lambdify() call still uses the original expr/symbols, so numerics are unchanged. The cache key separately carries the symbol-name tuple, so name-keying is safe and deterministic.

2. Replaces the flaky wall-clock tests in TestPerformanceExpectations (the original motivation). test_lambdify_caching's assert time2 <= time1*2 was failing CI on unrelated PRs (e.g. Refined-submesh (coarse/fine companion) investigation #192) from pure runner jitter, and its loose tolerance was masking bug uw.function.evaluate() never hits the lambdify cache (fresh sympy.Dummy per call → +1 cache entry every call) #194.

Tests

• test_lambdify_cache_hit — exercises get_cached_lambdified directly: identical request returns the same object, no new entry; distinct expression cached separately. Deterministic, timing-free.
• test_evaluate_cache_stable_across_calls — uw.function.evaluate() never hits the lambdify cache (fresh sympy.Dummy per call → +1 cache entry every call) #194 regression guard: repeated identical evaluate() holds _lambdify_cache size flat ([1,1,1,…], was [1,2,3,…]) and results identical across calls. Aggregate cache behaviour, no wall-clock.

Validation

• Cache size across 7 identical evaluate() calls: [1,1,1,1,1,1,1] (before fix: [1,2,3,4,5,6,7]).
• tests/test_0720_lambdify_optimization_paths.py — 19 passed, 1 pre-existing unrelated xpass.
• tests/test_0501_integrals.py smoke — 9 passed, 3 pre-existing xfail (unrelated CellWiseIntegral Perf: clone DM in Integral.evaluate() (3× speedup; closes #171 narrative) #172/Revert PR #172 (integrals regression) + xfail CellWiseIntegral parity tests #174). No numerical regression from the cache-key change.

Test plan

• TestPerformanceExpectations — deterministic, no timing
• Full test_0720 module — 19 passed
• test_0501_integrals safety smoke — 9 passed
• uw.function.evaluate() never hits the lambdify cache (fresh sympy.Dummy per call → +1 cache entry every call) #194 reproducer now flat
• CI green on this branch

Closes #194.

Underworld development team with AI support from Claude Code

[Copilot]

Pull request overview

This PR replaces flaky wall-clock assertions in lambdify optimization tests with deterministic behavioral checks for cache reuse and evaluation consistency.

Changes:

• Replaces timing-based lambdify cache test with a direct get_cached_lambdified cache-contract test.
• Replaces rbf=False timing assertion with a consistency check between rbf=True and rbf=False.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.