Performance audit: fix critical hotspots across extraction, storage, and retrieval

[LeahArmstrong]

Tier 1 (Critical):

• Cache estimated_tokens in ExtractedUnit (avoid repeated metadata.to_json)
• Use Sets for DependencyGraph type_index/reverse (O(1) vs O(n) include?)
• Memoize DependencyGraph#to_h (avoid 3x redundant serialization)
• Add store_batch to VectorStore interface (pgvector multi-row INSERT, Qdrant batch upsert)
• Reuse HTTP connections in Qdrant adapter (avoid TCP handshake per request)
• Wrap temporal snapshot inserts in transaction (single commit vs 5K fsyncs)

Tier 2 (High):

• Pre-compile regex constants in CallbackAnalyzer (avoid dynamic Regexp.new in loops)
• Eliminate flatten calls in orchestrator summaries (use each_value + sum)
• Reuse HTTP connections in OpenAI embedding provider
• Add find_batch to MetadataStore interface (single WHERE IN query)
• Batch metadata lookup in Ranker and ContextAssembler

Tier 3 (Medium):

• Cache @nodes.keys in PageRank loop
• Use max_by(limit) in GraphAnalyzer hubs (O(n) vs O(n log n))
• Create path Set once in git enrichment (not per batch)
• Reduce checkpoint frequency in embedding indexer (every 10 batches)

Tier 4 (Low):

• Combine AR_INTERNAL_METHOD_PATTERNS into single alternation regex

Also fixes: pre-existing Pathname require in shared_extractor_context,
view_template_extractor spec for root environments.

https://claude.ai/code/session_017z8KCJGSwnH3dScDiwsqov