cgen: add deref for aliased reference types using mut captures

[dy-tea]

Fixes #27355

[medvednikov]

@codex review

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[medvednikov]

Verdict: Yes, it's a good PR — correct, well-scoped, and idiomatic. ✅

What I verified

• Bug reproduces on master: the reproducer fails with error: struct or union expected / member reference base type 'main__AppRef'... is not a structure or union (one-deref-short access on the captured pointer).
• Fix works: with the PR applied and v rebuilt, the reproducer compiles and prints ok.
• No regressions: full vlib/v/tests/aliases/ suite passes (66 passed, 1 skipped, incl. the new test), plus targeted closure tests (closure_test.v, closure_of_method_defined_on_alias_test.v, mut_closure_fixed_array_thread_test.v, type_alias_of_fn_with_mut_args_test.v, alias_to_ptr_arg_test.v).
• Edge cases I added all behave correctly: pointer-alias to a non-struct (type IntRef = &int), read-only alias capture, plain mut struct capture, and passing an alias-typed variable (the else branch). Importantly, mutation propagates through the pointer (score = 2 after two calls), so the alias-as-pointer semantics are preserved, not silently copied.
• Formatting: v fmt -verify is clean on both files.

Why the approach is sound

The fix is three coordinated changes, and each one mirrors an existing pattern right next to it:

1. gen_anon_fn (struct init) and gen_anon_fn_decl (struct field type) — symmetrically extend the existing is_auto_deref && !var.is_mut deref so it also fires for mut captures when the type is an .alias. This makes the context-struct field a single pointer (AppRef = App*), so the body's ->score matches.
2. ref_or_deref_arg_ex — adds a branch parallel to the existing struct expected_type == arg_typ.ref() branch, emitting the extra & indirection for the mut alias arg. The new branch is tightly guarded (exp_sym.kind == .alias && unaliased_type(expected) == arg_typ) and is disjoint from the struct branch below it, so no shadowing.

Minor nitpicks (non-blocking)

• The test is weak: test_main() only checks that it compiles and runs — there's no assert a.score == 1. I confirmed the mutation does propagate, so adding an assertion would pass and would lock in the semantics, not just "doesn't crash." Worth requesting.
• Small asymmetry: the new PrefixExpr path does g.expr(...) without setting g.arg_no_auto_deref = true, whereas the analogous struct branch immediately below does. It's harmless for the &a case, but a reviewer might flag the inconsistency.
• Some structural duplication with the struct branch — acceptable for readability.

[medvednikov]

Thanks!