feat: rank parallels by textual relevance

[xr843]

Summary

• rank exact normalized sentences ahead of containing passages
• prefer shorter containing passages before MITRA confidence
• make final ties deterministic by source, juan, and Chinese text
• replace quadratic search-result grouping with HashMap lookup

Why

Common passages such as 色即是空 previously ranked long commentarial matches ahead of concise source passages because search ordering primarily used confidence and insertion order. This change makes relevance explicit and deterministic without changing the CLI, JSON schema, SQLite schema, or data artifact.

Verification

• cargo fmt --all --check
• cargo clippy --all-targets -- -D warnings
• cargo test --all (61 tests)
• python3 -m pytest tests/ -q
• cargo build --release
• real MITRA dataset: 色即是空 returns concise matches first in under 0.01 seconds
• real MITRA dataset: two-character 色即 query completes in 0.15 seconds