Fix review findings before 0.3.0rc3, and bump to 0.3.0rc3

[m-messer]

Fixes from a review of main since v0.3.0-rc.2, then bumps all four packages to 0.3.0rc3. One commit per fix.

expression_parsing

• A task alias of summation names the task's symbol again where it isn't applied: with S aliased to summation, 2*summation is 2*S (it raised "summation is a function"). Applied, summation(...) is still the sum.
• integrate, summation, product and limit are unevaluated without simplify (Integral, Sum, Product, Limit), like other functions. They are plain functions, so unevaluated parsing didn't reach them: product(k!, (k, 1, 3000)) and sum(k!, (k, 1, 3000)) took seconds and then failed.
• With simplify, SymPy works out products and limits and leaves integrals and finite sums unevaluated. check_evaluation_cost now bounds a Product with numeric limits: at most MAX_PRODUCT_TERMS (1000, exported) factors, with the index as large as its bound (product(k, (k, 1, 500)) is rejected like factorial(500)).
• A sum with an infinite limit is still summed (oo, pi**2/6): SymPy can't approximate a divergent series numerically, and compareExpressions compares by value.

criteria

• split_criteria raises CriteriaSyntaxError for an empty criterion (a = b,, a,,b, ""), instead of returning "" for the criteria parser to fail on.
• split_criteria treats all brackets as one kind, so half-open intervals split correctly: response in [0, 1), response > 0 is two criteria (it raised). A closing bracket with none open, or a bracket left open, still raises.
• Node shapes are defined once, on each node class's STYLE; graph_to_mermaid reads them like tree.py does. The output is unchanged.

slr_parsing

• A reduction with no goto raises GrammarError rather than a bare KeyError. Only an error action that rewrites the stack can cause one. _reduce and parsing_action go through ParseTable.goto/entry; parsing is about 4% slower on a criteria benchmark.

units

• unsplittable_names is cached per unit set (a small gain: about 1.6% of parse_quantity).

Version

• 0.3.0rc3 in every pyproject.toml, ranges between the packages >=0.3.0rc3,<0.4.0, README install tags v0.3.0-rc.3, CHANGELOG headings, poetry.lock.

Downstream (compareExpressions)

• Its test suite, run against this branch, passes except test_radians_to_frequency, which also fails on main (units: make the angle a dimension, so 1 Hz is 2π rad/s #1 addresses it).
• Behaviour changes it may notice: sums with finite limits, integrals, products and limits parse unevaluated without simplify. Its comparisons simplify the difference, which works them out, so equal answers still compare equal.
• It still needs a time limit around evaluation: .simplify() there works out expensive expressions.

Verification

• ruff, format and mypy are clean.
• All four test suites pass: slr_parsing 116, criteria 105, expression_parsing 515, units 538.
• The parity check shows 0 unexpected diffs out of 1608 probes.
• scripts/release_version.py v0.3.0-rc.3 --prerelease passes.
• New tests cover each fix.

Not in this PR

• With elementary_functions, integrate(...) is rewritten to int E grat E and fails to parse. This existed before this branch; it needs its own fix.

After merging, tag v0.3.0-rc.3 on main to run the pre-release workflow.

🤖 Generated with Claude Code