Summary

Full-scope collection inference can introduce a Salsa cycle when multiple unannotated collections are constrained by the same container literal:

from typing import Any

def run(cond: bool, d: dict[Any, Any]) -> list[Any]:
    a = {}
    b = {}
    if cond:
        b = d
    return [a.get("x", 0), b.get("x", 0)]

To infer a = {} and b = {}, we collect constraints from later uses such as a.get(...) and b.get(...). This creates a cycle because inferring the literals requires inferring the return statement, while inferring the calls requires resolving the literal definitions.

The conditional assignment makes b's collection-use constraint unstable across fixed-point iterations. Without that constraint, b remains dict[Unknown, Unknown], which exposes a single class specialization at b.get(...) and records the constraint on the next iteration. Once recorded, the literal refines to dict[Unknown, int | Unknown]; joining it with d produces dict[Unknown, int | Unknown] | dict[Any, Any], which no longer exposes a single class specialization, so the constraint disappears again. The constraint map therefore alternates between {a} and {a, b} even after the expression and binding types converge, eventually reaching Salsa's cycle-iteration limit and panicking. The existing cycle-initial values let evaluation begin but do not make this metadata monotonic.

Once the initial tainted iterations have passed, this preserves collection-use constraints from previous iterations across scope, definition, expression, and statement inference. The constraint map now grows monotonically and reaches a fixed point, while definition inference expands compact metadata only when needed and preserves its other retained fields. The regression coverage includes both direct returns and annotated assignments.

Closes astral-sh/ty#3778.