Improve generic localization quality and evidence preservation

[zzet]

Summary

• add a benchmark-independent localization quality specification with anti-overfitting and evaluation gates
• tighten declaration ownership for nested Go callables, PHP annotations, and C/Rust macros
• preserve declaration-backed evidence across refinement, exact-read, and recovery transitions
• keep terminal state honest when recovery is empty, exhausted, or evidence is shed
• protect concurrent localization reservations and add deterministic, synthetic regression coverage

Design constraints

• tree-sitter remains the primary source of syntax structure
• manual parsing is limited to bounded grammar, recovery, preprocessor, or external-block gaps
• production behavior is based on syntax, graph structure, query intent, and evidence provenance
• no new benchmark task IDs, repositories, paths, transcripts, or expected answers are embedded in implementation or tests

Verification

• go test ./internal/mcp ./internal/parser/languages -count=1
• go test -race ./internal/mcp -count=1
• go build ./internal/mcp/...
• formatting and diff checks
• independent review of ownership ambiguity, macro determinism, evidence preservation, and concurrent recovery behavior