feat(cache): Sprint 29 — streaming response cache with SSE replay (v1.9.0)

[konjoinfinity]

Summary

• New module konjoai/cache/streaming.py — StreamChunk, StreamingCacheEntry, StreamingResponseCache, singleton factory (get_streaming_cache)
• POST /query/stream now checks the streaming cache before running the retrieval pipeline; on a miss it records each token and inter-chunk timing, then stores the complete response after the stream ends
• On a cache hit, the route returns a StreamingResponse replaying stored SSE frames with cache_hit: true in every frame and configurable inter-chunk pacing (cache_stream_replay_delay_ms)
• 3 new config fields (all False/0 by default — K3):
  • cache_stream_enabled — master switch
  • cache_stream_replay_delay_ms — fixed sleep between replayed chunks (0 = no sleep)
  • cache_stream_max_chunks — safety cap; oversized responses are not stored
• 38 new tests covering models, lookup/store/replay, singleton factory, and route hit/miss/disabled paths

Invariant compliance

Gate	Status
K1 — no silent failures	store() wraps inner cache write in try/except; logs warning, never raises
K3 — graceful degradation	get_streaming_cache() returns None when either cache_enabled or cache_stream_enabled is False; route falls through to full pipeline
K4 — dtype contract	Dedicated SemanticCache inner instance asserts q_vec is float32 at store()
K5 — zero new hard deps	asyncio, json, threading, time, dataclasses — all stdlib; numpy already required
K6 — backward-compatible	Existing /query/stream SSE contract is unchanged on miss; cache_hit key absent on miss

Test plan

• python -m pytest tests/unit/test_streaming_cache.py -v — 38 tests pass
• python -m pytest tests/ -q — 1 165 passed, 7 skipped, 0 failures
• ruff check / ruff format --check — clean

https://claude.ai/code/session_0171rxj96JzbbH6JwuMu2zfx

---

Generated by Claude Code