[dgageot]

Why

The scheduler's worker lifetime was owned by separate Start/Stop calls,
which made lifecycle ownership (who starts, who stops, who owns restarts)
hard to reason about across nested tools and forked sessions, and left gaps
where completion/restart/probe signals could race or be lost.

What

• Replace Start/Stop worker ownership with a single blocking
  Run(ctx) on the scheduler; the worker's lifetime is now scoped to that
  call and tied to context cancellation.
• Introduce tools.Runnable as the canonical host bridge for this model, and
  migrate all callers: normal and additional toolsets, Code Mode, the
  fork-skill path, and raw session extras.
• Make team/runtime cancellation and join explicit instead of implicit via
  Stop.
• Fix nested lifecycle correctness: worker completion propagation, restart
  ownership, and probe non-blocking behavior, and distinguish lifecycle
  contention from actual signal loss.

Validation

• Independent iterative deep review — final verdict clean.
• CI=true task build, task test, task lint all pass.
• Affected-package race tests (tools, codemode, agent, runtime,
  teamloader, scheduler) repeated 10x.
• Cross builds verified for Windows amd64 and JS/WASM.
• SplitGoroutineLifetime audit via temporary overlay: 9 warnings, down
  from 10 (scheduler's removed); the checker remains disabled globally.

No further code changes intended in this PR.