feat(db): add PostgreSQL + pgvector storage adapter for semantic search

[yoshiiita]

Summary

Closes #145

Implements internal/db/pgvector.go — a production-grade PostgreSQL 16 + pgvector storage adapter for indexing code ASTs, diffs, and repository memory embeddings with sub-millisecond semantic search.

What's in this PR

• PGVectorStore — opens a Postgres connection and auto-migrates the schema on startup
• embeddings table — stores content with a vector(1536) column backed by an HNSW index (vector_cosine_ops) for approximate nearest-neighbour queries at sub-millisecond latency
• Upsert — idempotent insert-or-update of an embedding (keyed by content-derived ID)
• Search — top-k cosine similarity query, optionally scoped to a content kind ("ast", "diff", "memory")
• DeleteBySnapshot — bulk removal for GC / snapshot pruning
• New deps: github.com/pgvector/pgvector-go, github.com/lib/pq

Schema

CREATE TABLE embeddings (
    id          TEXT PRIMARY KEY,
    snapshot_id TEXT NOT NULL,
    kind        TEXT NOT NULL,   -- "ast" | "diff" | "memory"
    content     TEXT NOT NULL,
    embedding   vector(1536)
);

CREATE INDEX embeddings_hnsw_idx
    ON embeddings USING hnsw (embedding vector_cosine_ops);

Test plan

• Start a local Postgres 16 instance with pgvector enabled
• Call NewPGVectorStore(dsn) — verify extension + table + index are created
• Upsert several embeddings with different kind values
• Search with kind="" — confirm all kinds are returned in distance order
• Search with kind="ast" — confirm only AST rows are returned
• DeleteBySnapshot — confirm rows are removed
• go build ./... passes

[vercel]

@yoshita-madamala is attempting to deploy a commit to the Kavindu's projects Team on Vercel.

A member of the Team first needs to authorize it.