feat(prefill): add maintenance cache saturation harness

[FluffyAIcode]

Summary

• capture canonical first-append token IDs in a bounded in-memory queue, excluding maintenance replay sessions
• expose API-key-protected status/drain endpoints without persisting prompts or tokens
• add a staged replay driver for per-node 90%, 95%, and LRU churn targets with memory/fallback/publish safety stops
• expose cache evictions, bytes evicted, put failures, publish attempts/failures, and bytes published through capability cards, API, and dashboard
• emit privacy-safe JSON reports containing only salted capture IDs, counts, timings, and capacity metrics

Capacity semantics

The harness controls head 1 GiB and allens 8 GiB independently. Resident fleet usage cannot exceed 9 GiB; the churn stage verifies that excess logical writes increase eviction counters while resident bytes remain bounded.

Tests

• platform-neutral Python selection — 827 passed, 1 skipped
• focused capture/API/coordinator tests — 14 passed
• eviction/publish/capability/network tests — 39 passed
• TypeScript SDK — 54 passed, 100% coverage and typecheck
• proto lint, canonical format, generated stub drift, Python compile and diff checks pass

Maintenance execution

After merge, deploy with --cache-fill-capture-size 256, feed live gRPC first appends, run scripts/fill_prefill_cache_from_live_grpc.py, then validate staged fill, LRU churn, and remote replay before disabling capture.

Made with Cursor