[ty] Add bounded constraint solution projections

[carljm]

Constraint solution extraction can enumerate an exponentially large family of paths even when sequent derivation and intersection construction stay within their existing limits. This adds separate limits for solution traversal, collected paths, and retained type terms. Exceeding a limit aborts the projection without exposing a partial family as the complete answer.

ConstraintSet solution collection is fallible and bounded by default. Callers preserve the original signature or constraints, decline optional narrowing, or use Unknown when a projection cannot finish. Generic inference does not treat evidence from the remaining arguments as complete.

A fallible fold processes complete per-path bindings, preserving correlations between type variables until the consumer combines alternatives. Generic inference uses the collector and fold while retaining its merged-specialization behavior below the limits.

This is a foundation for astral-sh/ty#3557. It does not yet change call inference to intersect separately specialized return types.

Test plan

• Verify correlated alternatives and Cartesian products, exact budget boundaries, interning/source-order permutations, and early termination before or during solution selection.
• Exercise shared preprocessing/collection limits and restoration of path state after traversal and path-budget cancellation.
• Cover incomplete, rejected, and unsolved solutions; alias and set-theoretic term accounting; bounded DNF expansion; and conservative generic fallback.
• Add an mdtest with four independent intersection arguments, checking the existing inferred union in both argument orders.
• Preserve return-type inference from an overloaded callback with a catch-all overload and partially annotated literal-specific overloads.
• Verify Unknown recovery after genuine projection exhaustion, even with precise evidence from another argument, in both argument orders.

[astral-sh-bot]

Typing conformance results

No changes detected ✅

Current numbers

The percentage of diagnostics emitted that were expected errors held steady at 97.68%. The percentage of expected errors that received a diagnostic held steady at 93.44%. The number of fully passing files held steady at 109/136.

[astral-sh-bot]

Memory usage report

Memory usage unchanged ✅

[astral-sh-bot]

ecosystem-analyzer results

No diagnostic changes detected ✅

Full report with detailed diff (timing results)

[dcreager]

Looks good. Should we consider using this new trait for the fuel budget inside of PathAssignments at some point, as well? That would let it be a shared budget. Not at all a blocker.

[dcreager]

ah clippy

[carljm]

Looks good. Should we consider using this new trait for the fuel budget inside of PathAssignments at some point, as well? That would let it be a shared budget. Not at all a blocker.

Yeah I think that's worth evaluating. It's almost certainly possible to share more infrastructure; whether they should actually be shared budgets is less clear to me.