perf(rust): eliminate context cloning and merge duplicated evaluation paths

[aepfli]

Summary

• Take owned Value instead of &Value in all evaluate methods, avoiding a deep clone of the context on every evaluation. This is the biggest win for large contexts (1000+ attributes)
• Merge duplicated evaluation methods: evaluate_flag_internal + evaluate_flag_internal_pre_enriched collapsed into single evaluate_flag_core (-151 lines)
• Add merge_metadata_flag_set_only helper to avoid HashMap::new() allocation on flag-not-found path
• enrich_context now destructures owned Value instead of cloning the context map

Net: -148 lines (162 added, 310 removed)

Motivation

The evaluation hot path cloned the context Value on every call:

• Regular path: enrich_context called obj.clone() to deep-copy the entire context map
• Pre-enriched path: context.clone() passed to evaluate_owned even though nothing needed modification

Since evaluate_owned → Arc::new(data) takes ownership anyway, passing owned values from the start eliminates the clone entirely. The WASM callers already own the parsed context, so this is free.

The two nearly-identical evaluate_flag_internal / evaluate_flag_internal_pre_enriched methods (~130 lines each) were a maintenance hazard — any bug fix had to be applied in both. Merged into a single evaluate_flag_core with a needs_enrichment: bool parameter.

Test plan

• All 134 unit tests pass
• All 36 integration tests pass
• All 39 gherkin scenarios pass
• All 6 metadata merging tests pass
• cargo clippy -- -D warnings clean
• WASM target builds successfully
• Run benchmarks to measure improvement on large contexts

🤖 Generated with Claude Code