docs: correct architecture + metrics claims (Pinecone not pgvector, MCP-consumable, real numbers)

[DevanshuNEU]

Problem

Public-facing copy (README, landing) and derived claims contain inaccuracies that fail under scrutiny:

1. pgvector is claimed but NOT used -- grep for pgvector/vector( returns nothing; vectors live in Pinecone, Postgres holds metadata only (migrations/001_create_codeintel_schema.sql has no vector column).
2. "sub-300ms" -- no measured basis; remove until issue B produces real percentiles.
3. "used by Claude/Cursor" -- overclaim; true statement is "exposes an MCP server consumable by Claude Code / Cursor / Cline."
4. Store roles should be stated precisely: Pinecone = vectors, Supabase/Postgres = system of record (metadata), Redis = cache + rate-limit + pub/sub.

Acceptance criteria

• README/landing say Pinecone (not pgvector) and state each store's role precisely.
• No latency claim until backed by issue B.
• MCP phrased as capability ("consumable by"), not adoption.
• Embedding model stated accurately (voyage-code-3 @ 1024, fallback OpenAI text-embedding-3-small @ 1536).