Add ADR 0008: session-bound runtime + gRPC protocol; supersede ADR 0007

ADR 0008 records the architectural pivot from automatic server-side prefix
matching (ADR 0007) to an explicit session-bound runtime, driven by:

  * Empirical failure of ADR 0007's primitive against Qwen3's
    enable_thinking=False chat template (continuation_rate = 0.0 in
    local-only smoke #2 evidence).
  * User-stated 2026-06-01 strategic frame: standalone runtime for local
    agent applications, no framework coupling, no deadline, no sunk-cost
    reasoning, extreme KV efficiency, zero intelligence regression.
  * The need to make 'long-session usability' falsifiable on Mac M4
    without depending on any one tokenizer template's round-trip
    stability.

Key decisions captured:

  * gRPC bidirectional streaming as the primary wire protocol; HTTP+SSE
    preserved as a deprecated single-shot shim.
  * Server-issued session_id, raw-token-history, append-only contract.
    No chat template at the runtime ever.
  * Per-session sink+window KV slab, byte-exact bound to (session_id,
    history_token_ids). INV-1 / INV-2 / INV-3 refined for sessions.
  * Python and TypeScript SDKs as first-class v0.3 deliverables (TS
    targets Node/Electron/Bun, not browser).
  * Mac M4 integration test report as a binding pre-merge gate (\u00a79).

ADR 0007 is marked Superseded by 0008; none of its implementation
reaches main (recorded as decision C3=b on 2026-06-01).

Index README updated.

This PR ships only the markdown; implementation will land in PR
sequence A1..F1 documented in \u00a76. No code changes here.

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: cursoragent

docs/adr/0007-cross-request-kv-reuse.md

@@ -1,6 +1,6 @@
 # ADR 0007 — Cross-request KV cache reuse for long sessions
 
-- **Status**: Accepted (2026-05-31, blocking v0.3.0 GA)
+- **Status**: Superseded by [ADR 0008](0008-session-bound-runtime-and-grpc-protocol.md) (2026-06-01). Originally Accepted 2026-05-31; superseded after empirical falsification of §2.4 by Qwen3 chat-template re-rendering. None of this ADR's implementation reached `main`. See ADR 0008 §1.2.
 - **Date**: 2026-05-31
 - **Decision drivers**: ADR 0006 §2.3.b empirical evidence (4-hour Mac M4
   run); the gap between the "local agent infrastructure" framing and the