fix(pi-subagent): meaningful wait failures, 30m default timeout, resume after timeout

Three production-observed defects (156 tool errors in 14 days of live use):

1. Wait/final delivery threw the delivered text as the error. For
   output_mode:'file-only' runs the entire failure became 'Output written
   to /tmp/x.md'; for empty outputs '(no output)' — the actual failure
   reason was lost. Failures now lead with run state + per-task error
   summaries (composeRunFailureMessage), with the delivered payload
   trailing as context.

2. Default run timeout was 15m; three general-profile runs were killed
   mid 'git commit'/'gh pr create'. Default is now 30m, still
   configurable via ~/.pi/subagent.json defaultTimeoutMs /
   PI_SUBAGENT_TIMEOUT_MS. Explicit timeout_ms behavior unchanged.

3. Resume of a timed-out child failed in ~1s with only 'Subagent exited
   with code 1'. Root cause: pi-subagent handed pi a bare session id,
   which pi resolves via a cwd-filtered lookup that falls back to a
   global search with an interactive stdin prompt — fatal in rpc mode.
   Bare ids are now resolved to their session file under sessionDir
   before launch (missing file fails fast with an actionable error), and
   nonzero-exit failures now include the stderr tail (e.g. 'No session
   found matching ...') so future failures are diagnosable.