Adds ability to run long-running bash processes in the background without
blocking the AI conversation. Useful for builds, test suites, dev servers,
and other processes that take longer than typical tool timeouts.

Core components:
- CircularBuffer<T>: O(1) ring buffer for stdout/stderr (1000 lines each)
- BackgroundProcessManager: lifecycle management with spawn/list/terminate/cleanup
- Process IDs use 'bg-{8-hex-chars}' format for uniqueness

New tools:
- bash tool: enhanced with run_in_background parameter
- bash_background_read: check status and retrieve buffered output
- bash_background_list: discover processes after context loss
- bash_background_terminate: graceful shutdown (SIGTERM → SIGKILL)

Features:
- Automatic cleanup on workspace deletion
- Output buffers preserved after process exit
- Filtering support in read tool (tail, regex)
- Workspace isolation (processes scoped to workspace)

Test coverage: 84 tests across all components

Add supportsBackgroundExecution property to Runtime interface to detect
runtime capability. SSH workspaces return clear error message instead of
failing with ENOENT or running on wrong host.

TODO comments added for future SSH support via nohup/setsid.

The onExit callback was unconditionally setting status to 'exited',
overwriting the 'killed' status set by terminate(). Now guards with
if (process.status === 'running') to preserve killed/failed states.

Added test to verify killed status is preserved after termination.

When a process is killed by signal (SIGTERM, SIGKILL, etc.), Node's close
event provides code=null and signal name. Previously we defaulted null to 0,
making killed processes appear as successful exits.

Now converts signals to Unix-conventional exit codes (128 + signal_number):
- SIGKILL (9) → 137
- SIGTERM (15) → 143
- SIGINT (2) → 130

This allows bash_background_read/list to distinguish clean exits from
forced terminations (via terminate() or external kill/OOM).

… background execution

- Add BackgroundExecutor interface and BackgroundHandle for abstracting process spawning
- Create LocalBackgroundExecutor wrapping BashExecutionService for local processes
- Add SSHBackgroundExecutor stub (returns 'not yet implemented' error)
- BackgroundProcessManager now uses per-workspace executor registration
- Remove pid from public API (runtime-specific detail)
- Remove supportsBackgroundExecution from Runtime interface (executor error suffices)
- Share single BashExecutionService instance across all local executors

This refactor prepares for SSH background execution support by making the
execution backend swappable per workspace.

…start

- Add ensureExecutorRegistered() called from getOrCreateSession()
- Add Config.getWorkspaceMetadataSync() for sync metadata lookup
- Make registerExecutor() idempotent (no-op if already registered)

Without this fix, background execution tools failed for all pre-existing
workspaces after app restart with 'No executor registered' error.

Moves from pre-registration to lazy executor creation/caching:

- Remove registerExecutor()/unregisterExecutor() from BackgroundProcessManager
- spawn() now takes executor as first arg, caches per workspace
- AIService creates executor alongside runtime in streamMessage()
- Remove ensureExecutorRegistered() and getWorkspaceMetadataSync()
- Remove 3 registration calls from IpcMain workspace creation paths

This follows the existing pattern where runtime is created on-demand
from metadata rather than pre-registered. Executors are still cached
for SSH connection pooling.

Net ~100 LoC removed.

Remove BackgroundExecutor abstraction layer - Runtime already handles
local vs SSH execution differences, so background spawning belongs there.

- Add spawnBackground() to Runtime interface with BackgroundHandle type
- Implement LocalBackgroundHandle in LocalRuntime with event buffering
- Stub SSHRuntime.spawnBackground() (not supported for SSH workspaces)
- Simplify BackgroundProcessManager to call runtime directly
- Update bash tool to pass runtime from ToolConfiguration
- Remove aiService executor registration (no longer needed)
- Delete backgroundExecutor.ts, localBackgroundExecutor.ts, sshBackgroundExecutor.ts

~300 LoC removed, cleaner separation of concerns.

…cesses

When a background process is killed by signal (SIGTERM, SIGKILL, etc.),
Node's exit event provides code=null and signal name. The spawnBackground
handler was defaulting null to 0, making killed processes appear as
successful exits.

- Add getExitCode() helper to convert signals to Unix-conventional exit
  codes (128 + signal_number): SIGTERM=143, SIGKILL=137, etc.
- Add test verifying terminated processes report exit code >= 128

This restores behavior that was lost when bashExecutionService.ts was
removed during the Runtime refactor.

- Check both exitCode and signalCode when determining if process is alive
- Fixes terminate() always sending SIGKILL after SIGTERM, even when process
  already stopped (signal-killed processes have signalCode set but exitCode null)
- Also gate background tools for SSH runtimes (TODO: replace duck-type check
  with proper capability flag once SSH implements spawnBackground)

A single TextDecoder with stream:true buffers partial multibyte sequences.
When stdout and stderr interleave, shared state corrupts output by combining
bytes from different streams. Use independent decoders and flush on exit.

Adds terminateAll() to BackgroundProcessManager and hooks it into
Electron's before-quit event, preventing orphaned processes when
mux exits normally. Crash recovery (orphans from SIGKILL) remains
out of scope.