docs: label all performance claims as MEASURED vs PROJECTED (NO FAKE CODE compliance)

Fixed 10 evidence violations across 5 files per NO FAKE CODE policy:
- All billion-scale claims now labeled as PROJECTED (not yet benchmarked)
- Distinguishes calculated projections from empirical measurements
- Maintains architectural honesty about what's tested vs theoretical

Files updated:
- src/hnsw/typeAwareHNSWIndex.ts (2 claims)
- src/utils/metadataIndex.ts (1 claim)
- src/query/typeAwareQueryPlanner.ts (1 claim)
- docs/architecture/finite-type-system.md (2 claims)
- CHANGELOG.md (4 claims)

Changes:
- 87% HNSW memory reduction → PROJECTED (calculated from architecture)
- 86% metadata memory reduction → PROJECTED (calculated from chunking)
- 385x type tracking reduction → PROJECTED (calculated from Uint32Array)
- 40% query latency reduction → PROJECTED (calculated from graph reduction)

All claims remain architecturally sound but are now honestly labeled.
Future TIER 4 work will add benchmarks to upgrade PROJECTED → MEASURED.

Audit document: .strategy/EVIDENCE_VIOLATIONS_AUDIT.md

src/utils/metadataIndex.ts

@@ -124,7 +124,7 @@ export class MetadataIndexManager {
   // Reduces file count from 560k → 89 files (630x reduction)
   // ALL fields now use chunking - no more flat files
   // v3.44.1: Removed sparseIndices Map - now lazy-loaded via UnifiedCache only
-  // This reduces metadata memory from 35GB → 5GB @ 1B scale (86% reduction)
+  // PROJECTED: Reduces metadata memory from 35GB → 5GB @ 1B scale (86% reduction from chunking strategy, not yet benchmarked)
   private chunkManager: ChunkManager
   private chunkingStrategy: AdaptiveChunkingStrategy