feat: cache async dogfood results

[douglasmonsky]

Summary

• add warm result caching for async dogfood jobs keyed by request options plus local DB/config fingerprints
• cache compact aggregate dogfood payloads in memory and on disk under the dogfood cache directory
• expose result_cache metadata in usage_dogfood_status so agents can see hit/miss/source/cache key
• document the fast repeat pattern: run fresh once, then repeat with refresh=false and use_cache=true

Real local evidence

• cold usage_dogfood_start(refresh=False, evidence_limit=4, privacy_mode="strict", write_markdown=False, use_cache=True): ~12.2s
• next identical run after cache store: disk hit in ~0.0008s
• next identical run after memory load: memory hit in ~0.0002s

Validation

• PYTHONPATH=src .venv/bin/python -m pytest -q --tb=short
• .venv/bin/python -m ruff check .
• .venv/bin/python -m compileall src
• python3 scripts/check_release.py
• git diff --check

Notes

• Cache hits intentionally require refresh=False; refresh=True still refreshes/recomputes so agents do not accidentally hide newly indexed usage.
• Cached payloads are the existing compact aggregate dogfood artifacts, not raw logs or indexed snippets.