Fixes for roborev review 365:
- capture.rs: use length-prefixed binary format instead of R-side JSON
  construction. Rust handles all JSON serialization, eliminating
  escaping issues with control characters. Temp files use PID+counter
  for unique names (no more race conditions).
- server.rs: add CancellationToken + JoinHandle for graceful shutdown.
  :ipc stop now properly signals the server thread to exit, allowing
  :ipc start to work again.
- mod.rs: set r_is_at_prompt=false during evaluate to prevent
  concurrent evaluate requests from both thinking R is idle.
- server.rs: use PID-unique temp file for R version extraction.

Additional fixes:
- Add "ipc" to meta command autocomplete list (completion/meta.rs)
- Support :ipc stop then :ipc start by using Option<Receiver>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Use nchar(type = "bytes") consistently in R capture code guards
  (previously mixed nchar() and nchar(type = "bytes"))
- Remove misleading "(with timeout)" from stop_server join comment
- Add log message about in-flight connections on shutdown

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Use reedline's new ExternalBreak signal to interrupt read_line() when
IPC user_input is received. The injected code is echoed on the prompt
line, saved to history, and triggers spinner/duration tracking just
like normal user input.

- Add break_signal Arc<AtomicBool> to IPC module, shared with reedline
- Fire break signal when user_input stores pending input
- Handle Signal::ExternalBreak in both REPL loops (R mainloop + standalone)
- Switch reedline dependency to git rev with break signal support
- Add PTY integration test for IPC user_input flow

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Fix malformed HTTP body in IPC test (leading spaces from format! indentation)
- Add PID filtering to find_socket_path to avoid wrong session in parallel runs
- Add Windows CR strip to ExternalBreak handler for consistency with Success path

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Collapse nested if statements in pty_ipc_tests (collapsible_if)
- Suppress dead_code warning for socket_path on Windows
- Move sessions_dir import into cfg(unix) block

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Use tokio NamedPipeClient with tokio::time::timeout (300s) instead of
synchronous File I/O, matching the Unix socket timeout behavior.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

capture.output(print()) bypasses WriteConsoleEx, so with visible=true
the value never appeared in the REPL terminal. Now R code explicitly
cat()s the value and message()s the error when visible mode is active.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Add PTY-based tests for the evaluate IPC method:
- test_ipc_evaluate_capture: stdout via cat(), value, error, mixed
- test_ipc_evaluate_visible: visible=true shows output in REPL
- test_ipc_evaluate_silent: visible=false (default) hides from REPL

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Rework visible=true to act as a blocking version of send/user_input:
- Code is injected into the REPL prompt (appears like user-typed input)
- R evaluates normally with output shown in the terminal
- WriteConsoleEx callback captures output for the JSON response
- IPC client blocks until evaluation completes (R returns to prompt)

The previous approach evaluated silently in the idle callback and
printed results via cat(), but the code never appeared at the prompt.

Introduces PENDING_VISIBLE_EVAL state + check_visible_eval_completion()
which detects when R returns to the prompt and sends the deferred reply.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Review 378 (2d13303):
- Fix Windows compile: use cfg-gated eprint for strip_cr/format_error_output
- Release RwLock before terminal I/O in ipc_capture_callback
- Restore special chars test for 2-field capture file format

Review 380 (10d3add):
- Replace flaky sleep in test_ipc_evaluate_silent with sentinel pattern

Review 381 (1dd200b):
- Fix race: hold r_is_at_prompt=false during finish_ipc_capture
- Use unwrap_or_else for poisoned mutex in handle_visible_evaluate
- Add 300s staleness timeout for pending visible eval
- Use Acquire/Release ordering for at_prompt flag

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Upgrade Relaxed loads/stores to Acquire/Release in handle_request to
pair correctly with Release stores elsewhere. Document the timeout
path's best-effort race with active R evaluation.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Add ipc_tests.rs that verifies IPC JSON-RPC functionality without
depending on terminal output verification. Uses portable-pty for
process spawning with platform-specific transport (Unix sockets /
Windows named pipes), enabling IPC testing on Windows where PTY-based
terminal output capture is not available.

Tests: evaluate (value, stdout, error, mixed, visible, sequential),
user_input acceptance.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Set shutdown flag before sleep in Drop so reader thread can exit
  during grace period, then join the thread (best-effort)
- Document simplistic HTTP response parser relies on Connection: close
- Document PID=None cross-talk risk and --test-threads=1 recommendation
- Add module-level note about per-test process spawn cost

Co-Authored-By: Claude Opus 4.6 <[email protected]>

cargo fmt fixes and mark all IPC tests with cfg_attr(windows, ignore)
because ConPTY does not support crossterm cursor::position() — reedline
cannot initialize, so arf never reaches the prompt. True Windows IPC
testing requires a headless mode (no reedline).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Route all IPC operations through reedline's ExternalBreak mechanism
to check the editor buffer before accepting. This prevents two bugs:

1. IPC send/eval during console typing no longer corrupts the display
   - If the user has typed something, IPC returns USER_IS_TYPING error
2. Silent eval during typing is also blocked (previously froze the terminal)

Key changes:
- Replace IPC_PENDING_INPUT with PENDING_IPC_OPERATION enum
- Defer user_input reply until ExternalBreak confirms empty buffer
- Route silent evaluate through ExternalBreak instead of idle callback
- Show "agent> " prefix consistently for all IPC-injected code
- Show "[evaluating...]" indicator during silent evaluate
- Clear prompt line before IPC output to prevent display corruption
- Add USER_IS_TYPING (-32002) error code, remove unused INPUT_ALREADY_PENDING

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Deduplicate VisibleEvaluate/UserInput arms in ExternalBreak handler
  by extracting shared post-accept logic (clear prompt, history, spinner)
- Add explicit fall-through comment for SilentEvaluate in fast-path
- Use INPUT_ALREADY_PENDING (-32002) instead of R_BUSY when another
  IPC operation is queued, so clients can distinguish the two cases
- Document buffer.trim() semantics (whitespace-only treated as empty)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

When the REPL is in shell mode, history browser, or help browser,
the idle callback is not running and IPC requests would hang in the
mpsc queue until the 300s timeout. Add an IN_ALTERNATE_MODE flag
that is checked in both dispatch_request (server thread) and
handle_request (main thread) to immediately reject with
R_NOT_AT_PROMPT instead of waiting.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Consolidate alternate-mode tests into single test functions to avoid
flakiness from parallel execution sharing the global AtomicBool.
Add comments explaining why RAII guard is not needed for the
set_in_alternate_mode calls around browser invocations.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Update reedline to beb43685 which saves suspended_state in the
ExternalBreak handler instead of calling move_cursor_to_end(). This
prevents the extra blank line that appeared when re-entering read_line()
after a rejected or accepted IPC operation.

Add two PTY integration tests:
- test_ipc_user_input_reject_no_extra_line: verify prompt row stays
  the same after USER_IS_TYPING rejection
- test_ipc_user_input_no_extra_blank_line: verify no blank line between
  R output and the next prompt after IPC user_input

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

- Validate Content-Length against 1MB limit before reading body to
  prevent memory exhaustion via large Content-Length header
- Create session file with mode 0600 atomically via OpenOptions to
  eliminate TOCTOU permission gap
- Use case-insensitive Content-Length header matching per RFC 7230
- Hold SERVER_HANDLE mutex from check through write to eliminate
  TOCTOU race in start_server double-start guard

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

- Use checked_add for Content-Length size computation to prevent
  integer overflow from bypassing the 1MB request limit
- Document the known TOCTOU limitation for directory permissions
  (Rust std lacks mkdir-with-mode)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…ailure

- Use blocking lock instead of try_lock in stop_server visible eval
  cleanup to ensure capture callback is always removed and reply sent
- Log warning when sessions directory permission change fails

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

- Document intentional spec deviation for treating id:null as notification
- Restore panic safety rationale in alternate mode save/restore comments

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

id:null is a valid request id per JSON-RPC 2.0 spec and should receive
a response. Only skip response when id is truly absent.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…et directories

Replace create_dir_all + set_permissions two-step with DirBuilderExt::mode(0o700)
so the directory is created with correct permissions atomically. Applied to both
the sessions_dir and fallback /tmp paths.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…ion dir TOCTOU

- Replace expect() with early return of EvaluateResult error when temp
  file creation fails (e.g., disk full, permissions)
- Use DirBuilder::mode(0o700) in write_session for atomic directory
  creation, consistent with get_socket_path

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>