Preserve coherent evidence context and publish continuous-memory benchmarks

[momonpya]

Selected explanations could gain scattered fragments without completing their context, while repeated graph-key preparation made packet fitting expensive. This change preserves the original fitted packet and all admitted observations, refines only an already nominated explanation with additional coherent original context, and prepares immutable graph keys once. Lean contracts cover retained fragments, provenance, original-packet fallback, selection equivalence, and complete packet bounds.

The public benchmark now starts with empty memory and a newly recorded seed, saves through LLVM topics 1–3, and repeats those topics with accumulated memory. GPT-6 Luna xhigh used 76.0% fewer input + output tokens in the first lap and 96.4% fewer in the retained lap against the same three independent fresh references. Seed cost is separate; all runs, actual handoffs, answers, minor source-review qualifications, and performance measurements are included. Runtime LLVM tests and statistical quality equivalence are not claimed.

Validation: the full native engine suite, lifecycle regressions, setup/package tests, kernel contract audit, and all seven new native ephemeral model runs passed their execution gates. No incomplete initial handoff or timeout occurred; the slowest hook was 23.6 seconds within its 30-second deadline. All original observations and completed answers remained persisted with empty indexing queues before each next topic. The release workflow now publishes runtime.json alongside binaries and the plugin so every artifact describes the same build.