Label native crashes by type instead of always 'Stack overflow!'

docs/notes/setranslator-exception-labels.md

@@ -0,0 +1,78 @@
+# Working note — Fix seTranslator switch fall-through (mislabeled crash diagnostics)
+
+Branch: `fix/setranslator-exception-labels`
+Type: Bug / Reliability
+
+## Goal (restated)
+
+`seTranslator` (`src/blitzrc/bbruntime_dll/bbruntime_dll.cpp:55-70`) is the Windows
+structured-exception translator: it maps a CPU fault code to a human-readable
+panic message. Its `switch` has **no `break` statements**, so every case falls
+through and `panicStr` always ends as `"Stack overflow!"`. Result: every native
+crash — divide-by-zero, the very common null/dangling-object access violation,
+illegal instruction — is reported to the user as "Stack overflow!", sending them
+down the wrong debugging path. Directly contradicts INTENT #6 ("Trust the
+runtime… instrumented enough to diagnose"). Confirmed by recon ~5×; the top
+deferred Bug candidate.
+
+## Approach
+
+Add `break;` after each `case` in the switch so each exception code maps to its
+correct message. Minimal correctness fix — no new codes, no pipeline redesign.
+
+```cpp
+switch( u ){
+case EXCEPTION_INT_DIVIDE_BY_ZERO: panicStr = "Integer divide by zero"; break;
+case EXCEPTION_ACCESS_VIOLATION:   panicStr = "Memory access violation"; break;
+case EXCEPTION_ILLEGAL_INSTRUCTION:panicStr = "Illegal instruction"; break;
+case EXCEPTION_STACK_OVERFLOW:     panicStr = "Stack overflow!"; break;
+}
+```
+
+Message flow: `seTranslator → bbruntime_panic → RTEX → bbEx → debugError →
+Debugger::debugMsg`. In test mode (`-t`) the DummyDebugger prints `"Error: " + msg`
+to stdout and `exit(1)`; in normal runs it shows the message box.
+
+## Verification
+
+`-t` mode routes the panic message to stdout (DummyDebugger), so it is capturable:
+a `.bb` that triggers a runtime integer divide-by-zero (divisor in a variable to
+avoid constant folding), run with `blitzcc -t`, should print
+`Error: Integer divide by zero` (post-fix) vs `Error: Stack overflow!` (pre-fix).
+Plan: confirm this empirically against the rebuilt binary; if capturable, add a
+CLI-contract check to `test.bat`/`test.sh` using a crash fixture generated OUTSIDE
+`tests/` (a panic calls `exit(1)`, so it must NOT be a normal `-t` suite file —
+that would abort the whole run). If the harness can't cleanly capture it, fall
+back to documented manual repro + the obvious correctness of the four `break`s.
+
+## Files touched
+
+- `src/blitzrc/bbruntime_dll/bbruntime_dll.cpp` — `seTranslator` (4 `break`s).
+- `test.bat` / `test.sh` — CLI crash-diagnostic contract (if capturable).
+- `docs/notes/setranslator-exception-labels.md` — this note.
+
+Windows-only by nature (`_set_se_translator`); macOS has its own signal path
+(out of scope). No ABI/codegen/parser change.
+
+## Acceptance criteria
+
+- A runtime integer divide-by-zero reports "Integer divide by zero", not
+  "Stack overflow!" (verified empirically; ideally pinned by a CLI check).
+- Unknown codes still report "Unknown runtime exception".
+- `test.bat` green; build 0 errors; no regression.
+
+## Trade-offs / rejected
+
+- **Skipped: adding more exception codes** (FLT_DIVIDE_BY_ZERO, PRIV_INSTRUCTION) —
+  scope creep beyond the missing-break bug; keep one-purpose.
+- A normal `-t` Test block can't assert this (the panic `exit(1)`s mid-run) — hence
+  the standalone-fixture CLI-contract approach.
+
+## Deferred follow-ups (high priority, surfaced this iteration)
+
+- **GC doc false citation (my own #67 error):** `help/language/lang_ref_modern.html`
+  cites `tests/GarbageCollectionTest.bb` as proof of the "reference cycles leak —
+  break them yourself" contract, but that test has NO cycle case. Either add the
+  cycle/aliasing test cases (GC acceptance suite — a recon pitch this iteration) or
+  correct the citation. Worth doing soon (shipped-doc correctness).
+- Contextual-keyword Phase 2 (`Release` + expression-position keywords).

scripts/fixtures/divzero.bb

@@ -0,0 +1,19 @@
+Strict
+EnableGC
+
+; Deliberate runtime integer divide-by-zero. Used by test.bat / test.sh to verify
+; the Windows structured-exception translator (seTranslator in bbruntime_dll.cpp)
+; labels the fault correctly as "Integer divide by zero" -- not the pre-fix
+; mislabel "Stack overflow!" caused by missing `break`s in its switch.
+;
+; This program PANICS and exits non-zero by design, so it lives OUTSIDE tests/
+; (it must not run in the `blitzcc -t` suite loop) and outside samples/tutorials/
+; games/ (the corpus compile-sweep). The harness invokes it explicitly with -t and
+; asserts the panic message.
+;
+; The divisor is a variable so the divide is not constant-folded at compile time.
+Test divideByZeroFault()
+	Local denom = 0
+	Local result = 10 / denom
+	Assert( result = 0 )
+End Test

src/blitzrc/bbruntime_dll/bbruntime_dll.cpp

@@ -58,12 +58,16 @@ static void _cdecl seTranslator( unsigned int u,EXCEPTION_POINTERS* pExp ){
 	switch( u ){
 	case EXCEPTION_INT_DIVIDE_BY_ZERO:
 		panicStr = "Integer divide by zero";
+		break;
 	case EXCEPTION_ACCESS_VIOLATION:
 		panicStr = "Memory access violation";
+		break;
 	case EXCEPTION_ILLEGAL_INSTRUCTION:
 		panicStr = "Illegal instruction";
+		break;
 	case EXCEPTION_STACK_OVERFLOW:
 		panicStr = "Stack overflow!";
+		break;
 	}
 
 	bbruntime_panic( panicStr.c_str() );

test.bat

@@ -23,6 +23,10 @@ call :expect_failure "EOF error names end of file, not a control byte" "end of f
 call :expect_ide_format "IDE machine format preserved under blitzide" "!DIAG_BB!"
 del /q "!DIAG_BB!" >nul 2>&1
 
+rem --- runtime crash diagnostic: native faults are labeled by type, not all "Stack overflow!" ---
+rem The fixture panics and exits non-zero by design, so it lives outside tests\.
+call :expect_failure "native divide-by-zero is labeled correctly" "Integer divide by zero" -t "%ROOTDIR%\scripts\fixtures\divzero.bb"
+
 cd /d "%ROOTDIR%\tests"
 
 for /R %%f in (*.bb) do (

test.sh

@@ -169,6 +169,26 @@ if grep -q "error:" "${TEST_TMPDIR}/diag-ide.out"; then
   FAILED=1
 fi
 
+# --- runtime crash diagnostic: native faults are labeled by type ---
+# The seTranslator must map each structured exception to its own message rather
+# than reporting every fault as "Stack overflow!". The fixture deliberately
+# integer-divides by zero; in -t mode the panic prints "Error: <msg>" and exits
+# non-zero, so it lives outside tests/ and is invoked explicitly. (Windows-only
+# path; on hosts without the structured-exception translator this is a no-op
+# check that simply won't see the message -- so only assert it where it applies.)
+CRASH_BB="${ROOTDIR}/scripts/fixtures/divzero.bb"
+if [[ -f "${CRASH_BB}" ]]; then
+  set +e
+  "${BLITZCC}" "${TARGET_FLAG[@]}" "${EXEC_FLAG[@]}" -t "${CRASH_BB}" > "${TEST_TMPDIR}/crash.out" 2>&1
+  crash_rc=$?
+  set -e
+  if grep -q "Stack overflow" "${TEST_TMPDIR}/crash.out" && ! grep -q "Integer divide by zero" "${TEST_TMPDIR}/crash.out"; then
+    echo "crash diagnostic FAILED: divide-by-zero mislabeled as Stack overflow" >&2
+    cat "${TEST_TMPDIR}/crash.out" >&2
+    FAILED=1
+  fi
+fi
+
 while IFS= read -r -d '' f; do
   if ! "${BLITZCC}" "${TARGET_FLAG[@]}" "${EXEC_FLAG[@]}" -t "${f}"; then
     echo "\"${f}\" failed at least one test"