Dedup tparam-default substitution between alt and solver layers

[nitishagar]

Summary

The logic for substituting references to earlier type parameters inside a tparam's default was duplicated in two places:

• get_tparam_default in pyrefly/lib/alt/class/targs.rs (used during call/specialization)
• an inlined copy in finish_class_targs in pyrefly/lib/solver/solver.rs (used during class-level finalization), flagged by a // TODO: deal with code duplication in get_tparam_default marker.

The two copies differed only in their lookup source and their fallback value for out-of-scope references — differences that are intentional, since one path runs where an out-of-scope reference is an already-reported error (returns Any(Implicit)) and the other runs during finalization with no such error (returns the param's gradual type).

This PR extracts a shared substitute_tparam_defaults helper into pyrefly_types::quantified (next to the existing as_gradual_type_helper, which already uses the same TypeVar/TypeVarTuple/ParamSpec/Quantified match idiom). The helper takes the default by value, a &dyn Fn(&Name) -> Option<Type> lookup closure, and a caller-chosen fallback Type, so each call site keeps its own lookup source and fallback semantics. Both callers are rewritten to delegate to it, and the // TODO marker is removed.

Behavior

This is a behavior-preserving refactor. Each call site passes the same lookup and fallback it used inline, so the produced types are identical for every input:

• get_tparam_default passes name_to_idx -> checked_targs lookup with self.heap.mk_any_implicit() fallback (unchanged).
• finish_class_targs passes seen_params -> new_targs-then-targs lookup with param.as_gradual_type() fallback (unchanged).

The two fallbacks are deliberately not collapsed — they intentionally differ (error path vs finalization path).

Test plan

• cargo build -p pyrefly_types — clean
• cargo build -p pyrefly — clean
• cargo test -p pyrefly --lib — 7675 passed, 0 failed, 3 ignored (the full typechecker regression suite; zero behavior diff)
• python3 test.py --no-test --no-tensor-shapes --no-conformance --no-jsonschema — clean for the touched files

Fixes #4298

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D114523923. (Because this pull request was imported automatically, there will not be any future comments.)

[rchen152]

I'm not convinced this is an improvement. Helpers add conceptual overhead, and callbacks add implementation complexity, whereas the original code is simple and clear, and the duplication is small.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[nitishagar]

Thanks for taking the time. A couple of thoughts on the trade-off, happy to drop this if it's not convincing:

• The shape of this helper (default + lookup closure + fallback) is what TODO: deal with code duplication in get_tparam_default #4298 proposes, and the // TODO marker it points at has been sitting in finish_class_targs for a while. So the dedup is fulfilling a tracked request rather than being introduced speculatively.
• My main motivation is drift risk: the two copies do the same TypeVar/TypeVarTuple/ParamSpec/Quantified match, and a future change to one (e.g. handling a new tparam form) would have to be remembered in both. The two intentionally differ only in lookup source and fallback, which are exactly the two parameterized pieces.

That said, I take the point on the &dyn Fn adding implementation weight. If the conceptual overhead isn't worth it for this size of duplication, I can just close this and #4298. Let me know which way you'd like to go.