fix(hy-emit): fill an elided literal's head through parentheses (fixes #630)

[Leimiu-G]

Fixes #630.

What

hy-emit/src/expr.rs::emit_expr_exp is the only place that fills an elided composite literal's head
from its surrounding type. It matched the expression directly, so a redundant parenthesis put the
literal out of reach of the ty: None / elem: None arms, expected was discarded, and the emit
fell through to the "spell the type" refusal — on a program hyb check passes with 0 diagnostics.

m := map[string]svc{"a": ({replicas: 2, tag: "x"})}   // check: OK   build: error [emit]
m := map[string]svc{"a": {replicas: 2, tag: "x"}}     // check: OK   build: OK

Peel Paren before the expected-type dispatch:

 let mut lit = e;
 while let Expr::Paren { inner, .. } = lit { lit = inner; }
 match lit {
     Expr::StructLit { ty: None, fields, .. } => { /* fill head */ }
     Expr::SliceLit  { elem: None, items, .. } => { /* fill head */ }

Paren is documented as span-only (hy-syntax/src/ast.rs: "kept for faithful spans"), so the
elision has to apply identically through it. Dropping the parens from output is safe at all three
call sites — each is an element slot of a composite literal, where Go gives parentheses no meaning.
The peel is local to emit_expr_exp on purpose: Self::peel_parens is private to stmt.rs today,
and #619 promotes it separately, so this PR keeps to one hunk in one file.

Verification

Emitted Go is byte-identical with and without the parens, and the program output matches:

$ hyb build elided2.hy -o out && ./out      # map[a:{2 x}]
$ hyb build elided1.hy -o out && ./out      # map[a:{2 x}]

Tests — two new pins next to the existing diff_elided_map_value_struct_literal:

• diff_elided_map_value_struct_literal_under_parens
• diff_elided_slice_elem_struct_literal_under_parens (also covers double parens)

Falsifiable: with expr.rs reverted and the tests kept, exactly those two fail and the other
171 pass — no collateral, no test that passes without the fix.

with fix:      test result: ok.     173 passed; 0 failed
without fix:   test result: FAILED. 171 passed; 2 failed   # the two new ones

Workspace and differential corpus on 28754ab (patched):

• cargo test --workspace — all suites ok, 0 failed
• cargo test -p hy-emit --test '*' — ok
• tools/differential_loop.py — accept-side 21/23, reject-side 56/56

The 2 accept-side misses are the local environment, not this change: both are a7-* cgo fixtures
that fail assembling under a non-ASCII %TEMP% (can't create C:\Users\????\AppData\Local\Temp\ go-build...\_x006.o). They fail identically on unpatched main.

Related

Same invariant as #618 (parenthesized receiver loses its type) and #609 (line-leading ( swallowed
as an argument list) — the 8th independent locus where a site dispatches on Expr shape without
peeling first.

[limitlessprojects]

Landed as two-parent merge b8e850b onto main (worktree merge of pr-631 @ de0a97a, GitHub marks this PR merged). Thank you @Leimiu-G — clean, minimal, and exactly the right fix.

Live-bypass probe on main @ 55cd957 (before landing): confirmed LIVE — check-accepts/emit-rejects gap:

scenario	check	build
map[string]svc{"a": ({replicas: 2, tag: "x"})}	OK (0 diagnostics)	REJECT type-elided struct literal — spell the type
[]svc{({replicas: 2}), (({replicas: 3}))}	OK	REJECT same
controls without parens (map value / slice elem)	OK	built OK

Post-merge verification on b8e850b: both paren scenarios now emit and build (service{replicas: 2, tag: "prod"} etc.); unparenthesized controls unchanged.

Battery on merged tree: cargo test --workspace 1386/0 (+2 = the PR's two tests), corpus 100/100 verdict agreement (tripwire quiet), differential FULL 34+66.

Closes #630. Credit: Leimiu-G.