feat: improve oracle curation with deep mode, multi-query, and rev-pinning

- Default mode changed from keyword to "deep":
    patternType:keyword (NLS semantic matching) instead of plain keyword search
- Multi-query aggregation: generates N query variations from the search_pattern
  by OR-splitting terms and mining CamelCase identifiers from seed_prompt
- Files must appear in >= min_hits queries (default 2 of 3) to enter the oracle
  This requires an agent to replay multiple independent queries to recover all
  oracle files — reduces single-query tool-affinity bias
- Rev-pinning: every SG query gets `rev:<ref>` from fixture's `revision` field
  (tags, short hashes, HEAD all handled; mirrors already pinned but now explicit)
- Graceful degradation: effective_min = min(min_hits, len(variations)) so
  single-term patterns (ClusterRole, NotFound) don't over-filter
- New CLI flags: --mode [deep|nls|keyword], --num-runs N, --min-hits K
- Metadata: discovery_method="sourcegraph_nls_multiquery" distinguishes new
  oracles from legacy "sourcegraph_api" ones

Note: LLM-powered Deep Search (SG Cody) is not accessible over HTTP;
use the MCP batch agent approach for that capability.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sjarmak