feat: Build incremental AST caching layer to eliminate redundant re-parsing

[Nanle-code]

Overview

Every invocation of scan() re-parses all target .sol files from scratch, even when files have not changed since the last scan. In the VS Code extension use case (scan-on-save), this means the entire workspace is re-parsed on every keystroke-triggered save, which is wasteful and adds perceptible latency on large projects.

Problem

For a project with 50 Solidity files, a full workspace scan re-parses all 50 files even if only 1 changed. The @solidity-parser/parser is fast, but at scale (e.g. a large DeFi protocol with 100+ contracts and inheritance chains), this becomes the dominant cost in the scan pipeline.

Measured impact in benchmarks:

• 10 files: ~40ms parse time (acceptable)
• 50 files: ~200ms parse time (noticeable)
• 100+ files: ~500ms+ parse time (degrades UX)

Proposed Solution

Implement a content-addressed AST cache with the following design:

interface ASTCacheEntry {
  contentHash: string;      // SHA-256 of file contents
  ast: ASTNode;
  parsedAt: number;         // timestamp
  filePath: string;
}

1. Before parsing, compute SHA-256 of file content
2. Check an in-memory LRU cache (bounded to 200 entries) for a matching hash
3. On cache hit, return cached AST — skip @solidity-parser/parser entirely
4. On cache miss, parse and store in cache
5. In the VS Code extension, persist cache to globalStorageUri across sessions
6. Expose a clearCache() export from @chainproof/core for testing and forced refresh

Performance Target

• Re-scan of unchanged workspace: >80% reduction in parse time
• Single changed file scan: cache hit for all other files, only changed file re-parsed

Acceptance Criteria

• ASTCache class implemented in packages/core/src/ast/cache.ts
• LRU eviction policy with configurable max size (default: 200 entries)
• VS Code extension persists cache to context.globalStorageUri
• CLI does not persist cache between runs (in-memory only)
• clearCache() exported from @chainproof/core
• Benchmark results documented in PR description showing parse time reduction