fix: propagate output types through named nodes and sequences

[zharinov]

Summary

• Fix type mismatch when named definitions returning structured types (enums, structs) are used without capture inside named nodes
• Add new diagnostics for ambiguous cases: AmbiguousUncapturedOutputs and UncapturedOutputWithCaptures
• Extract compute_merged_flow helper to deduplicate logic between infer_named_node and infer_seq_expr
• Fix typegen to emit entrypoints with array result types

Why

When extracting a pattern into a named definition and referencing it, the behavior should be preserved. Previously:

Q = (program [X: (x) Y: (y)])  ; produces enum

But after extraction:

A = [X: (x) Y: (y)]
Q = (program (A))              ; produced void (bug!)

Now both produce the enum type correctly. Arrays also propagate: (program (A)+) produces A[].

Notes

• "Scalar" terminology replaced with "output" for clarity (scalars typically mean primitives, but we're dealing with structured types)
• Errors are intentionally strict to help LLMs self-correct rather than silently fail