…ments (#21 reviewer pass)

Reviewer-pass on PR #17 (v1.6 #21 LLDB suspend_thread landing). Three
findings, all addressed:

1. CRITICAL: dispatcher mapped backend errors to -32001 vs -32004 by
   substring-matching `e.what()` for "not implemented". Any backend
   that threw a descriptive error containing those words for a
   legitimate runtime failure would have been silently promoted to
   kNotImplemented. Fix: typed `backend::NotImplementedError` subclass
   of `backend::Error`; dispatcher catches the subclass first
   (→ -32001), generic Error second (→ -32004). GdbMiBackend's
   suspend_thread throws the new type. Catch order matters and is
   pinned in a comment — the compiler does not warn on reversed
   order.

2. Stale comment blocks describing thread.suspend as "still returns
   -32001 kNotImplemented for LLDB-backed targets" — that's the
   pre-#21 contract. Replaced both: the describe.endpoints text and
   the function-level comment in handle_thread_suspend. The function
   comment now enumerates the three dispatch branches in order (RSP
   channel → vCont;t, LLDB-backed → backend.suspend_thread,
   NotImplemented-throwing backend → -32001).

3. MINOR: stale `[legacy]` Catch2 tag on a test that was updated to
   assert the new non-RSP forwarding path. Renamed to `[non-rsp]`.

New tests (test_dispatcher_nonstop.cpp) lock in the typed-error
contract:
- `NotImplementedError → -32001 by exception type` — message
  deliberately does NOT contain "not implemented" to prove the
  dispatcher uses the exception type, not its message.
- `generic Error with 'not implemented' substring stays -32004` —
  inverse case; this is the exact regression the pre-fix code was
  vulnerable to.

Also updated the gdbmi `todo()` helper (currently `[[maybe_unused]]`)
to throw NotImplementedError, so the contract is uniform for future
stubs.

70/70 ctest green. +2 unit cases.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…etion)

Drops the -32001 kNotImplemented stub that thread.suspend returned for
LLDB-backed targets and lands a real per-thread suspend.

Why this works without SetAsync(true): SBThread::Suspend() is a flag
flip — the suspend bit gates the NEXT SBProcess::Continue, not the
suspend call itself. The async-mode flip is only required for the
*delivery* of stop events (the listener path), not for marking a thread
as parked. RSP-backed targets keep their existing vCont;t route (#17
phase-2); LLDB-backed targets now flow through the new backend virtual.

- DebuggerBackend grows a `suspend_thread(target_id, tid)` virtual.
  LldbBackend resolves SBThread by id and calls Suspend(); throws
  backend::Error on bad target/tid or no live process. GdbMiBackend
  throws "not implemented" — MI's per-thread semantics differ enough
  to warrant a separate item.
- Dispatcher::handle_thread_suspend forwards to backend_->suspend_thread
  for non-RSP targets. The "not implemented" branch surfaces -32001 so
  agents still see kNotImplemented when the backend genuinely lacks
  support; everything else is -32004 kBackendError.
- New test_backend_suspend_thread.cpp (4 cases) exercises live LLDB-
  launched suspend against the sleeper fixture + error paths.
- test_dispatcher_nonstop.cpp's "-32001 kNotImplemented in phase-1"
  case is replaced with two cases: forwards-to-backend (happy path)
  and unknown-target → -32004. test_dispatcher_vcont_rsp.cpp's "non-
  RSP returns -32001" case is updated to assert the new "no process"
  -32004 against an empty target.
- All 7 stub-backend test files grow a one-line suspend_thread
  override.

Stretch (SetAsync(true) flip + pump_until_stopped wrappers around
continue/step) intentionally deferred — docs/27 §7 already calls out
that cascade as its own commit. Ships minimum scope clean: 70/70 ctest
green, +5 unit cases, no other behaviour change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ments (#21 reviewer pass)

Reviewer-pass on PR #17 (v1.6 #21 LLDB suspend_thread landing). Three
findings, all addressed:

1. CRITICAL: dispatcher mapped backend errors to -32001 vs -32004 by
   substring-matching `e.what()` for "not implemented". Any backend
   that threw a descriptive error containing those words for a
   legitimate runtime failure would have been silently promoted to
   kNotImplemented. Fix: typed `backend::NotImplementedError` subclass
   of `backend::Error`; dispatcher catches the subclass first
   (→ -32001), generic Error second (→ -32004). GdbMiBackend's
   suspend_thread throws the new type. Catch order matters and is
   pinned in a comment — the compiler does not warn on reversed
   order.

2. Stale comment blocks describing thread.suspend as "still returns
   -32001 kNotImplemented for LLDB-backed targets" — that's the
   pre-#21 contract. Replaced both: the describe.endpoints text and
   the function-level comment in handle_thread_suspend. The function
   comment now enumerates the three dispatch branches in order (RSP
   channel → vCont;t, LLDB-backed → backend.suspend_thread,
   NotImplemented-throwing backend → -32001).

3. MINOR: stale `[legacy]` Catch2 tag on a test that was updated to
   assert the new non-RSP forwarding path. Renamed to `[non-rsp]`.

New tests (test_dispatcher_nonstop.cpp) lock in the typed-error
contract:
- `NotImplementedError → -32001 by exception type` — message
  deliberately does NOT contain "not implemented" to prove the
  dispatcher uses the exception type, not its message.
- `generic Error with 'not implemented' substring stays -32004` —
  inverse case; this is the exact regression the pre-fix code was
  vulnerable to.

Also updated the gdbmi `todo()` helper (currently `[[maybe_unused]]`)
to throw NotImplementedError, so the contract is uniform for future
stubs.

70/70 ctest green. +2 unit cases.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>