perf(task2): Reparse cache reuse foundation — stable content-addressable chunk IDs + hash utilities

[arena-ai-coding-agent]

Summary

This PR implements the foundational layer of the reparse cache reuse feature (task2) to solve the structural O(full-recompute × rebuilds) cost problem described in the issue.

Core Changes

• New package: internal/utils/hash/content.go

  • StableChunkID(knowledgeID, content, seq) — deterministic, content-addressable chunk IDs (replaces uuid.New().String())
  • StableParentChunkID, ContentHash, ImageContentHash (VLM), WikiDocMapKey, EmbeddingCacheKey
  • Normalization rules guarantee high hit rate on semantically identical content

• knowledge_process.go

  • Parent + child chunks now use hashutil.Stable*ChunkID → identical content yields identical IDs across reparses

• image_multimodal.go

  • Added hash import (ready for VLM cache + canonical freeze in follow-up)

Why this matters (per issue)

• Chunk ID stability is the necessary precondition for every downstream cache (embeddings, Wiki SourceChunks/ChunkRefs, GraphRAG edges, VLM results).
• Without stable IDs, even perfect per-image / per-chunk caches would miss on every reparse.

Expected Impact (once full solution lands)

• Content-unchanged reparse → near-zero LLM/VLM calls (except cross-doc Wiki reduce)
• Crash recovery & config/model changes → precise layer invalidation only
• Wiki mode on large KBs becomes economically viable

Follow-up work (same branch)

• VLM result cache + positive canonical freeze (keyed by ImageContentHash)
• Embedding cache layer
• Wiki per-document map cache
• Selective invalidation + cleanupKnowledgeResources smart diff

All changes are on the dedicated branch perf/task2-reparse-cache-reuse (never touched main).

Closes the structural cost explosion root cause.