fix(lldb): hide the Windows CRT, NT and winpthreads frames the stack filter never knew (#824)

[debugmcpdev]

Summary

Fixes #824.

On Windows the LLDB display filter hid nothing. Two causes, both in packages/shared/src/interfaces/lldb-policy-shared.ts:

1. The source rule did not recognise CodeLLDB's @symbol label. A frame without debug info arrives as source: {name: '@NtWaitForSingleObject', sourceReference: N} with no path, and FrameAnchorResolver carries that label into file. isSystemOrMissingSource('@NtWaitForSingleObject') returned false (not empty, not <unknown_source>, no POSIX path), so even a name the filter knew would not have been hidden — adding Windows names alone would have changed nothing.
2. No Windows names. The filter knew __libc_start_main, _start and the POSIX syscall wrappers, not __tmainCRTStartup, mainCRTStartup, BaseThreadInitThunk, RtlUserThreadStart, NtWaitForSingleObject, WaitForSingleObjectEx, DbgBreakPoint/DbgUiRemoteBreakin (the break-in thread a pause lands on) or winpthreads.

Change

• isSystemOrMissingSource: an @-prefixed file is "no source"; \ is normalised to / before path matching; MSYS2's mingw64|mingw32|ucrt64|clang64|clangarm64/include/, the MSVC vctools tree and Windows Kits are system paths.
• WINDOWS_RUNTIME_NAME_PATTERNS: the CRT start-up ([w]mainCRTStartup, [w]WinMainCRTStartup, __tmainCRTStartup, __scrt_common_main[_seh], invoke_main, __acrt_*), the kernel32/ntdll plumbing (BaseThreadInitThunk, RtlUserThreadStart, Dbg{BreakPoint,UiRemoteBreakin}, Nt*/Zw*, WaitFor{Single,Multiple}Object[s][Ex], Sleep[Ex], SleepConditionVariable{CS,SRW}, WaitOnAddress) and _?pthread_* — all under the existing name + no-user-source rule, so a user function named Sleep with a workspace path is kept.
• Nothing changes in FrameAnchorResolver: the [QUESTION] Go stack filtering can return an empty stack — should it keep the first frame like JS? #346 guarantee (never an empty stack) and the js: when the paused frame is inside node_modules, get_stack_trace hides it and get_local_variables/evaluate_expression silently anchor on an async ancestor frame #672 paused-frame disclosure apply as before.
• Docs: docs/stack-trace-filtering.md (the Windows rule, the @symbol shape, COBOL composes the filter too), docs/cpp/README.md step 6; changelog.d/824.fixed.md (it states the Linux side effect: pthread_* now also hides glibc's nptl frames, whose ./nptl/ path the filter already classified as system).

Measured on this box (Windows 11, CodeLLDB 1.11.8, MSYS2 g++ 15.2.0, examples/cpp/pause_test.cpp paused in its sleep loop, through the dev server)

Program thread, before → after:

	frames shown	hiddenFrames
before	12 (NtWaitForSingleObject … main … RtlUserThreadStart), none hidden	—
after	2: std::this_thread::sleep_for<…> (MinGW this_thread_sleep.h), main (pause_test.cpp:20)	10

includeInternals: true shows the full 12 with the 9 @symbol frames flagged unresolvedSource: true and the #816 note, as before.

The break-in thread the pause itself lands on (DbgBreakPoint → RtlUserThreadStart, 4 frames) now reads "All 4 frames are internal/runtime frames; showing the top internal frame …" with hiddenFrames: 3 — the #346 shape. That a launched program's pause_execution anchors get_stack_trace on this injected thread rather than the program's thread is a separate finding, filed on its own.

Tests

• adapter-policy-cpp.test.ts: 17 new rows in the real Windows shape (name: 'NtWaitForSingleObject', file: '@NtWaitForSingleObject' …), including the keep cases (user Sleep with a workspace path, main, sleep_for from the MinGW C++ header, a stripped user frame whose @label matches no name), plus filterStackFrames reduces the measured Windows pause stack to the program over the issue's 12-frame stack → [sleep_for, main].
• adapter-policy-rust.test.ts: 11 new rows (Windows names, MSVC invoke_main under vctools, glibc nptl pthread_cond_wait under ./nptl/ → hidden, user pthread_helper → kept, a /rustc/ std frame with a non-wrapper name → kept).
• adapter-policy-cobol.test.ts: three Windows rows (COBOL wraps the shared filter).
• session-manager-dap.test.ts (get_stack_trace gives the js-debug sourceMaps hint for native frames on a C/C++ (CodeLLDB) session #816 fixture): the default view now hides the @symbol frames (hiddenFrameCount ≥ 2, pausedFrame names NtWaitForSingleObject) and includeInternals: true still carries the unresolvedSource flags.
• tests/e2e/mcp-server-smoke-cpp-attach.test.ts: on win32 asserts hiddenFrames > 0 and a hidden note instead of the conditional unresolvedSource check that would otherwise silently stop running. Run locally: 2/2 pass.

Verification: npm run typecheck:all (ratchet 0/0), npm run lint, node scripts/check-docs.mjs, pnpm changelog:check; targeted unit run 379/379; pre-push hook (unit + integration) passed.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!