Surface pricing-feed degradation that silently weakens the routing budget

Follow-up from #982.

The routing budget is only as good as the pricing feed, and two degradation modes are currently invisible to operators:

1. Empty rates map. If the cost feed (models.dev / DO catalog) is unreachable at startup or refresh, the rates map is empty and every switch fails open (reason="no_pricing"). Startup is bounded by client timeouts (added in #982), but the only signal is a fetch log line — there is no metric to alert on.
2. Cached-rate fallback. When a feed doesn't publish a cached-read rate for a model, Plano silently estimates cached_rate = input_rate x cache_read_discount (default 0.1). If the provider's real discount differs, switch costs and the never-switch baseline are mispriced with no indication. (The DO catalog now publishes cache_read_input_price_per_million, so this mainly affects models.dev entries and DO models missing the field.)

Proposed:

• A gauge for cost-feed health (models loaded, last successful fetch timestamp) suitable for alerting.
• A counter or per-decision span attribute marking when the switch-cost calculation used the fallback discount rather than a published cached rate.
• Docs: call out both modes in the routing_budget config reference.

Related idea worth considering here: a static/local pricing override (JSON/YAML file merged over the feed) to remove the hard runtime dependency on an external pricing service — useful for self-hosted models whose IDs no public feed knows.