…generate output

Implements RouterStrategy::Cascade in zeph-llm (closes #1339).

When `strategy = "cascade"` is configured, the router tries providers in
chain order and escalates to the next only when the response is classified
as degenerate (empty, repetitive, incoherent, truncated).

Key behaviors:
- Heuristic classifier (default): zero LLM calls, detects degenerate outputs
  only (not semantic failures) — documented accurately
- Judge mode: LLM-based quality scoring with automatic fallback to heuristic
- Network/API errors do not consume escalation budget
- Best-seen response returned on budget exhaustion (not NoProviders)
- max_cascade_tokens caps cumulative token cost across escalation levels
- Skipped for chat_with_tools (falls through to standard routing)
- Thompson/EMA distributions not contaminated by quality-based failures
- quality_threshold validated at config boundary (clamped to [0.0, 1.0])
- window_size clamped to minimum 1
- collect_stream capped at 1 MiB to prevent unbounded allocation
- Token estimation uses chars().count() for correct non-ASCII handling

Config: [llm.router.cascade] with quality_threshold (default 0.5),
max_escalations (default 2), classifier_mode, window_size, max_cascade_tokens.

…zard, tests

- REV-01: record_availability(false) on stream-open errors in cascade_chat_stream
  early-provider loop, symmetric with cascade_chat error handling
- REV-02: add doc comment to RouterConfig::chain field
- REV-03: add round-trip serde test for RouterStrategyConfig::Cascade
- REV-04: add Cascade option to --init wizard step_router with quality_threshold
  and max_escalations prompts; wire through WizardState and build_config