fix(native): emitError chained method call caused segv on every compile error

[cs01]

User impact

The native self-hosted compiler (`.build/chad`) segfaulted instead of printing any compile error whose upstream code path ran through `emitError`. Concretely: fuzz A2-A5 repros (`new Date()` at module scope, `instanceof` with class hierarchy, `typeof d.getTime()`, etc.) all crashed with no output. Users saw a silent SIGSEGV where they should have seen a diagnostic.

After this PR, every one of those inputs prints the clean error the Node compiler already prints:

```
error: cannot determine type of module-scope variable 'd' (expression type: new). Move the declaration inside a function, or add a type annotation
```

Root cause

`emitError` in `src/codegen/llvm-generator.ts` built the output via a chained expression:

```typescript
this.diagnostics.formatDiagnostic(
this.diagnostics.getErrors()[this.diagnostics.getErrors().length - 1],
)
```

This is `method_call()[method_call().length - 1]` — an index access into the result of a method call. The native compiler's codegen for this pattern does not propagate the result's `elementInterfaceName` through, so the indexed element is handed to `formatDiagnostic` as a bare `i8*` pointing to memory the reader treats as a `%Diagnostic` struct with a different layout than what was stored. Subsequent `diag.message` GEP loads garbage; `strlen` on the garbage pointer crashes.

Confirmed under lldb: backtrace lands in `strlen` called from `formatDiagnostic + 260` (the `bold(": " + diag.message)` expression), with x1 holding ANSI-escape byte values instead of a string pointer. Debug prints before and after the intermediate-variable refactor prove the fix: with intermediates, `severity` reads `0` and `message` reads the correct string; without, severity reads `9.48606e-322` (garbage) and the message read segfaults.

Change

• `src/codegen/llvm-generator.ts` — `emitError` extracts `getErrors()` result and the last diagnostic into intermediate variables before passing to `formatDiagnostic`. Functionally identical; mechanically avoids the chained-index codegen bug.
• `tests/fixtures/edge-cases/new-date-module-scope.ts`, `tests/fixtures/edge-cases/new-date-zero-module-scope.ts` — regression fixtures using `@test-compile-error` annotation. They used to segv; now they emit the expected diagnostic.

Follow-up

The underlying codegen bug (method-call-result indexed access loses element interface type) is still present and worth fixing at the root in `access/member.ts` / `handleIndexAccessPropertyAccess`. This PR is the targeted workaround to immediately unblock A2-A5. Filed as future work.

Test plan

• `npm run verify` green (tests + stage 1 + stage 2)
• /tmp/chad_fuzz/c29_date, c36_date_basic, c37_date_zero, c38_date_typeof, c62_instanceof — all now emit clean compile errors (no segfault)
• Two regression fixtures added to `tests/fixtures/edge-cases/`

[github-actions]

Benchmark Results (Linux x86-64)

Benchmark	C	ChadScript	Go	Node	Place
Binary Trees	1.374s	1.258s	2.725s	1.207s	🥈
Cold Start	1.6ms	0.8ms	1.2ms	27.3ms	🥇
Fibonacci	0.814s	0.763s	1.559s	3.179s	🥇
File I/O	0.119s	0.093s	0.087s	0.207s	🥈
JSON Parse/Stringify	0.004s	0.005s	0.019s	0.015s	🥈
Matrix Multiply	0.454s	0.598s	0.573s	0.360s	#4
Monte Carlo Pi	0.389s	0.410s	0.405s	2.250s	🥉
N-Body Simulation	1.671s	2.125s	2.199s	2.381s	🥈
Quicksort	0.215s	0.247s	0.212s	0.262s	🥉
SQLite	0.354s	0.376s	—	0.433s	🥈
Sieve of Eratosthenes	0.016s	0.029s	0.020s	0.039s	🥉
String Manipulation	0.008s	0.018s	0.017s	0.037s	🥉

CLI Tool Benchmarks

Benchmark	ChadScript	grep	node	xxd	Place
Hex Dump	0.435s	—	0.910s	0.132s	🥈
Recursive Grep	0.019s	0.010s	0.098s	—	🥈