Document machine-readable responses and repair job scheduling

[crimson-knight]

What changed

• adds md / markdown response helpers and keeps the canonical emitted media type during negotiation
• documents exact application file placement for response formats and background jobs
• makes custom job retry backoff policies actually run after deserialization
• connects idle-only job workers to Amber's live HTTP request count while excluding upgraded WebSockets
• documents strict queue ordering, the unbounded process-local memory adapter, work-stealing limits, and channel broadcasts after jobs complete

Why

The public V2 guides need to describe behavior that the framework actually implements. The retry hook was documented but ignored, and work stealing had a request counter that the request pipeline never updated.

Developer impact

Applications can return Markdown through respond_with, custom retry schedules are honored, and idle-only workers now use a real request-activity signal. Existing queue and response behavior remains the default.

Validation

• focused responder, worker, and pipeline specs: 46 examples, 0 failures
• complete Amber V2 suite: 2,324 examples, 0 failures
• Crystal 1.21.0