Two coupled fixes shipped together. Both center on the daemon's shutdown
behaviour: the first stops orphaning agent subprocesses, the second is the
groundwork that lets future git operations actually be cancellable.

## 1. Daemon shutdown safety for the agent goroutine

`runAgentCmd` was creating its own `context.WithTimeout(context.Background(),
10*time.Minute)` and was launched bare (`go s.runAgentCmd(...)`), unparented
from the daemon's signal-handled context. On SIGINT/SIGTERM/idle-timeout the
agent subprocess (Setsid via daemonSysProcAttr) kept running, burning LLM API
tokens until its 10-minute budget elapsed, and then called AddReply /
RefreshFileList / RefreshDiffs / notify on a session whose watcher and SSE
goroutines were already gone — after WriteFiles had run.

Changes:
- `Server` gains `shutdownCtx` and `bgWG` fields plus `SetShutdownCtx`,
  `effectiveCtx`, and `WaitBackground` helpers.
- `runServe` calls `srv.SetShutdownCtx(ctx)` right after wiring the signal
  context, so all server-spawned background work parents on the same ctx.
- `handleAgentRequest` wraps the goroutine in `bgWG.Add/Done`.
- `runAgentCmd` derives its 10-min timeout from `effectiveCtx()` instead of
  `context.Background()`, and skips the post-run RefreshFileList/RefreshDiffs
  fan-out if the shutdown ctx is already done (the reply is still added to the
  in-memory session and persisted by the subsequent WriteFiles).
- The shutdown path in `runServe` calls `srv.WaitBackground(30*time.Second)`
  between `session.Shutdown()` and `session.WriteFiles()` so in-flight agent
  replies land before the file is rewritten. 30s ceiling so a wedged agent
  can't hang the daemon indefinitely.
- On `initErr`, `runServe` now calls `stop()` immediately instead of
  `<-ctx.Done()`. Without this the daemon would sit in 503-land for up to an
  hour after a failed init, holding a port and a process slot.

Tests: added `TestRunAgentCmd_CancelledByShutdownCtx` (verifies shutdown ctx
cancel kills the subprocess promptly) and `TestWaitBackground_TimesOut`
(verifies the wait helper drains cleanly and times out as expected).

## 2. `runGit` helper

`git.go` had ~42 `exec.Command("git", ...)` call sites, each repeating the
`exec.Command` + `cmd.Dir` + `cmd.Output()` + error-wrap pattern, with no
shared place to add cross-cutting concerns (ctx cancellation, env hardening,
stderr capture).

Added `runGit(ctx, dir, args...) ([]byte, error)` that uses
`exec.CommandContext`, sets dir, appends `GIT_TERMINAL_PROMPT=0` to the env
(prevents the daemon from blocking on a credential prompt with no tty), and
captures stderr in the returned error.

Converted the high-traffic call sites: `IsGitRepo`, `RepoRoot`,
`AllTrackedFiles` (used by handleFilesList), `changedFilesStaged`,
`changedFilesUnstaged`, `changedFilesBranch` (used by ChangedFilesScoped via
the session refresh path). Migration is intentionally partial — the helper
docstring notes that other call sites should be converted opportunistically.

The pre-existing test helper named `runGit` (in testutil_test.go) was renamed
to `gitT` — purely mechanical, all _test.go updated in one pass.