AbstractCore v2.13.21

Added

• ADR baseline: added an accepted ADR set covering engineering guardrails, validation/evidence, provider and capability ownership boundaries, server trust boundaries, source-first fixes, and the planned text-generation adapter lifecycle contract.
• Prompt-cache research backlog: recorded narrower proposed backlog items for exact-prefix memory-cluster recipes, external cache-binding semantics, and transformers/GGUF parity boundaries so future cache work starts from current code reality instead of stale assumptions.

Changed

• Loaded-runtime execution model: gateway-loaded local runtimes now route prompt-cache control-plane calls, bloc-KV ensure/load operations, and chat generation through one stable provider worker thread, while streaming responses bridge out through an unbounded queue instead of tying provider progress to client drain speed.
• Planning and docs ownership: the package backlog now points to docs/backlog/overview.md as the canonical planning entry point, the old ad hoc docs/KnowledgeBase.md was retired in favor of ADR-backed durable policy, and the adapter example docs now describe the current vLLM lifecycle reality without claiming portable model= hot-switching.

Fixed

• Loaded-runtime thread affinity: resident local runtimes no longer risk wedging when streaming cleanup happens on a different ASGI worker thread, and thread-affine local prompt-cache/bloc reuse paths now stay on the same provider thread across save/load/update/generate operations.