Self-host: gen-c writes defer as an unsupported comment and the deferred call never runs #7352

Profile (read first)

You are working as Gamma (Compiler Core) on epic #5980: rewrite t27c in t27 itself.
The self-hosted core (specs/compiler/core/t27core.t27) must write exactly the bytes
t27c gen-c writes, so every gen-c defect is a defect the core has to refuse instead of learn.
Rules for this task:

• Change gen-c only inside the Boundary below. gen/ is generated: never hand-edit it.
• Build with cargo build --release -p t27c and use ./target/release/t27c from the repository
  root, not a t27c on PATH.
• Write the regression test first and watch it fail on the unfixed compiler. Then fix, and
  revert the fix once to watch the test go red again. Say in the PR that you did both.
• Any spec that stops compiling must be listed in the PR by path with its fix. A silent corpus
  change is a failed task.
• The PR body says Closes # plus this issue's number and Part of #5980. ASCII only in source files.

Context

gen-c does not lower defer. A defer CALL; statement, the only form the corpus uses (13 specs,
e.g. defer BogatyrRegistry_deinit(&registry);), is written to C as the comment
/* unsupported: StmtExpr */; and the deferred call never runs. Exit status is 0 and the C
compiles, so a test that depends on the deferred call traps at run time, or worse, passes while
the cleanup it claims never happened. defer { ... } is a parse error ("Unexpected token in
expression: LBrace").

Zig runs deferred statements at every exit of the enclosing block (fall-through, return,
break, continue) in reverse order of declaration. A lossless lowering must do the same; if
that is out of scope, gen-c must refuse defer with an error instead of writing a comment.
The core (P1 slice 3) refuses defer with E_SYNTAX.

Reproduce

cargo build --release -p t27c
printf 'module m;\nvar n: i64 = 0;\nfn inc() -> void { n = n + 1; }\nfn f() -> void {\n    defer inc();\n}\ntest t { f(); assert(n == 1); }\n' > /tmp/d.t27
./target/release/t27c gen-c /tmp/d.t27 > /tmp/d.c; echo "gen-c exit $?"
grep -c unsupported /tmp/d.c
cc -std=c99 -DT27_TEST_MAIN /tmp/d.c -o /tmp/d && /tmp/d; echo "run exit $?"

Today: gen-c exit 0, 1, and the run aborts on assert(n == 1).

Acceptance criteria

• 1. Either the Reproduce run exits 0 with no unsupported comment in the C, or gen-c exits non-zero naming defer and its line
• 2. If defer is lowered: a fn with defer a(); defer b(); runs b before a, and a deferred call runs on an early return, on break and on continue inside a while body
• 3. cargo test --release -p t27c --test genc_defer passes

Boundary

• bootstrap/src/compiler.rs
• bootstrap/tests/genc_defer.rs

User Scenarios & Testing

• Given a fn whose first statement is defer inc(); and which returns early,
  when ./target/release/t27c gen-c lowers it and the C runs,
  then inc has run exactly once on each path, or gen-c refused the spec.

Requirements

• FR-001: gen-c MUST NOT write /* unsupported: ... */ in place of a statement and exit 0.
• FR-002: a lowered defer MUST run at every exit of its block, in reverse order.
• FR-900: cargo test --release -p t27c --test genc_defer MUST pass, and MUST fail with the fix reverted.
• FR-901: cargo test --release -p t27c --test core_selfhost and the corpus suite MUST still pass.

Part of #5980