Add deterministic code and snippet memory identity

[hunterbastian]

Summary

Implements deterministic identity metadata for code annotations and personal snippets so XMem can avoid re-judging exact code/snippet memories with an LLM.

Changes:

• add stable Pinecone metadata helpers for snippet identity, snippet search text, code annotation identity keys, and code annotation content hashes
• route code and snippet memory through deterministic judge paths using metadata lookups
• store snippet_hash, annotation_key, and annotation_hash in Pinecone metadata
• keep snippet code exact in metadata while embedding only the searchable description/language/tags text
• add regression coverage for repeated snippets across sessions and same-target code annotation updates

This addresses the edge case discussed in #141 where a user sends a snippet, then asks for the same snippet in another session. The normalized snippet_hash lets the judge no-op the duplicate without another model call.

Verification

• python3 -m compileall src/schemas/code.py src/agents/judge.py src/pipelines/ingest.py src/pipelines/weaver.py tests/unit/test_schemas.py tests/test_deterministic_memory_layer.py
• uv run --extra dev pytest tests/unit/test_schemas.py tests/test_deterministic_memory_layer.py -> 12 passed
• uv run --extra dev pytest -> 44 passed
• uv run ruff check --select F401 src/schemas/code.py src/agents/judge.py src/pipelines/ingest.py src/pipelines/weaver.py tests/unit/test_schemas.py tests/test_deterministic_memory_layer.py
• git diff --check

/claim #141