v0.6.4 — production_mode: retry-aware cost

Highlights

• production_mode: { fallback: "..." } on compare_models / optimize_retry_policy — measures retry-aware, end-to-end cost per successful output. Each candidate runs with a runtime-injected [candidate, fallback] retry chain.
• New metrics: escalation_rate, single_shot_cost, effective_cost, single_shot_latency_ms, effective_latency_ms, latency_percentiles — on both Report and AggregatedReport (averaged across runs:).
• Extended ModelComparison#table: Chain / single-shot / escalation / effective cost / latency / score. Edge case candidate == fallback → em-dash (not 0%), retry injection skipped so effective == single-shot by construction.
• context[:retry_policy_override] — new context key for transient per-call retry-policy overrides without mutating the step class.

Scope

• Single-fallback (2-tier) chains only.
• Step-only: raises ArgumentError if used on Pipeline::Base subclasses (pipeline-wide fallback semantics are a separate design question).

Docs

• Guide: Production-mode cost measurement
• CHANGELOG