A single Tolerance abstraction for rtol/atol #308

Problem / use case

#289 put the rtol/atol check into one function, values_within_tolerance in app/context/tolerance.py. The way a tolerance is chosen is still spread out:

1. Two sources of truth. EvaluationParameters.rtol/atol and parsing_parameters["rtol"/"atol"]. _tolerances in physical_quantity/graphs/match.py reads one and checks the other.
2. "Not set" and 0 look the same. rtol: 0 can't switch off the implicit significant-figures tolerance for physical quantities: 1.04 m vs 1.0 m is graded correct with rtol: 0.
3. The implicit tolerance applies only to physical quantities. 1.04 m vs 1.0 m passes, but symbolic 1.04 vs 1.0 fails.
4. The sig-fig formula is duplicated, in the app and in packages expression_parsing/numbers.py (compute_relative_tolerance_from_significant_decimals).
5. Loose floats are passed through evaluation.py (_apply_tolerance_aliases, _apply_precision_parameters, _build_parameters) and match.py (_tolerances, _within_tolerance).

Proposed solution

A frozen Tolerance dataclass, built once in _build_parameters and passed to both contexts:

@dataclass(frozen=True)
class Tolerance:
    rtol: float | None = None   # None = not set
    atol: float | None = None
    source: str = "default"     # "params", "sig_figs", "decimal_places", "implicit"

    def contains(self, res, ans, unit_factor=1) -> bool: ...

• None means unset, so rtol: 0 really means exact.
• Decide whether the implicit sig-fig tolerance applies to symbolic answers too, or to neither context.
• Use the package's sig-fig helper and remove the app copy.
• contains replaces values_within_tolerance, and does the unit scaling now in _within_tolerance.

Alternatives considered

Keep the floats and add a sentinel for "unset". That is smaller, but it leaves points 1, 3 and 5 as they are.

Related: #289, #300, #270.

Jira: LAM-323

🤖 Generated with Claude Code