feat(dispatch): implement Phase 16 strategy benchmarking

[404khai]

Phase 16 replays the same graph, orders, rider fleet, readiness seed, and traffic inputs through four independent dispatch policies: nearest feasible rider, lowest pickup ETA, readiness-aware lowest completion time, and preparation-aware assignment. The new roadrunner benchmark dispatch <scenario.json> [--idle-penalty-weight <weight>] [--json] command reports observed delivery durations, pickup waiting and rider idle time, distance, lateness, utilization, and explicit unassigned/unfinished populations.

All strategies share eligibility, road feasibility, deterministic RiderId tie breaking, decision evidence, and atomic assignment. Completion-time ranking includes readiness with zero waiting penalty; existing basic retains its road-travel objective. Nearest ranking records its meter-valued distance separately from informational travel seconds.

The collector verifies exact paired replay and standalone equivalence, then records hardware, dataset/graph identities, configuration, compiler/binary hash, raw process timings and median/p95/p99. Two committed synthetic comparisons use eleven measured runs per policy. They demonstrate different proximity/ETA and preparation-penalty behavior without claiming universal strategy superiority or production scalability. Methodology and completion evidence are in docs/dispatch-strategy-benchmarks.md and docs/phase-16-completion.md.

Validation:

• Formatting and strict workspace Clippy pass.
• cargo +1.99.0 test --workspace --all-features --locked: 138 tests pass; doc tests pass.
• Release CLI build and both replay-checking benchmark collections pass.
• Local documentation links and git diff --check pass.