Fix false dispatch timeouts on live JSONL activity

[amittell]

Summary

• include session JSONL mtime and pending/terminal JSONL state in dispatch status auto-resolution
• keep running labels from being marked interrupted when sessions.json is stale but JSONL shows active or terminal work
• separate watcher result probes from hard stall failures so high-thinking jobs are probed early but not failed until the idle-failure window
• deliver structured completion at watcher deadline before treating a freshly auto-interrupted status as failure

Live-path verification

• confirmed the scheduler daemon runs the installed package at ~/.openclaw/packages/openclaw-scheduler/node_modules/openclaw-scheduler/dispatcher.js
• confirmed live chilisaus delivery jobs execute ~/.openclaw/chilisaus/index.mjs and watcher.mjs as branded dispatch glue
• synced the same dispatch index/watcher fix into the live package copy and chilisaus wrapper copy; no wrapper-only logic changes were needed

Validation

• npm test (using deployed runtime dependency tree; 1768 passed, 0 failed)
• node --test tests/delivery-fixes.test.mjs (14 passed, 0 failed)
• npm run lint (using existing dev dependency tree because npm ci under Node 26 cannot build better-sqlite3@11.10.0 on main)
• npm run typecheck

Notes

• npm ci on this host's default Node 26 fails before tests because better-sqlite3@11.10.0 does not compile against Node 26/V8 14.2 in current main. This PR does not change dependencies; validation used the already-deployed runtime dependencies for tests and an existing dev dependency tree for lint/typecheck.

[amittell]

Superseded by #14, which re-applied this JSONL liveness fix onto current main together with the rest of the valuable branch/worktree fixes. Keeping the source branch for now; main contains the code via e21fdeb.