Previous benchmarks simulated core operations with inline loops that
didn't exercise the actual code paths. Now each benchmark calls the
real implementation:
- build_index_real: full build_index() on petstore fixture
- search_real_pet: SearchEngine::search("pet") against real index
- search_real_multi_term: multi-term query "list all pets"
- search_real_case_insensitive: case-insensitive search path
- normalize_json_input: normalize_to_json on JSON input
- pipeline_normalize_new vs pipeline_normalize_old: proves the
double-parse elimination (new returns tuple vs old parse-twice)
- detect_format_json_no_hints: byte-prefix format sniffing
- detect_format_with_content_type: format detection with hint
All benchmarks now use black_box correctly (consume return values,
not just inputs) to prevent dead-code elimination.
