[m-messer]

Problem

The symbolic and physical-quantity contexts each had their own rtol/atol check, and they disagreed (bug report, bug 9). This PR is stacked on #286 and replaces #262 (fix/atol-rtol-alignment), which made the same change on the code from before the refactor.

	Symbolic (compare_numerically)	Physical quantity (_within_tolerance)
Both rtol and atol set	both had to pass	only rtol was checked, unless the answer was 0
Boundary	<	<
Values that can't be converted to float	not within tolerance	raised

Changes

The agreed rules are the same as numpy.isclose, and they are what #262 proposed: |ans - res| <= atol + rtol*|ans|.

• values_within_tolerance (new app/context/tolerance.py):
  • applies that rule;
  • if symbols cancel only in (ans - res)/ans (e.g. both sides are multiples of rho*L**3), it checks rtol alone, because atol needs a numerical value of the answer;
  • returns False when the values can't be compared numerically.
• compare_numerically (app/context/symbolic/numerical.py) and _within_tolerance (app/context/physical_quantity/graphs/match.py) both call it.
  • For physical quantities, atol is still multiplied by the answer's unit factor, so it stays in the answer's units.
• app/docs/user.md: the tolerance descriptions now say ≤ and explain how rtol and atol combine. Also fixes some broken LaTeX in the absolute_tolerance entry.

Behaviour changes (answer 1 unless noted):

Case	Before	Now
rtol 0.1, atol 0.5, response 1.55	rejected (both contexts)	accepted
atol 0.5, response 1.5 (exactly at the tolerance)	rejected	accepted
answer rho*L**3, response 1.05*rho*L**3, rtol 0.1, atol 0.01	rejected	accepted

A physical quantity whose tolerance comes from the answer's significant figures also accepts a value exactly at the boundary now (1 m accepts 1.5 m).

Checklist

• Tests added/updated
• Docs updated (app/docs.md) if user-facing behaviour changed

Testing

• New tests:
  • test_numerical_tolerances in symbolic_evaluation_test.py and in physical_quantity_evaluation_test.py. The five cases that change behaviour failed before the fix.
  • test_values_within_tolerance calls the helper directly, including zero, complex and symbolic values.
• pytest app: 1290 passed, 2 skipped, 1 failed. The failure is test_radians_to_frequency, which already failed before this PR.
• mypy app: no issues.

With #284, #286, #288 and this PR, every bug in the report has a fix.

🤖 Generated with Claude Code