Speed up recall and upgrade knowledge building to GPT-6 Luna

[cucupac]

Recall now runs the existing MiniLM model through ONNX and fetches concept/evidence bundles in batches. This removes PyTorch/Sentence Transformers startup and repeated graph reads while retaining synthesis and retrieval limits. Upgrades prepare and verify pinned model files; recall loads them lazily and stays offline.

Measured on the production code across eight queries × three paired repeats:

Local recall	Median	Observed p95
Main	3.05s	3.24s
This PR	0.81s	0.88s

Production won 24/24 pairs, with no failures or timeouts. All pairs retained the same ordered memories and complete synthesis content; only anchor order changed to become deterministic. These timings include process startup but exclude external synthesis and background sync.

Latency validation: 672 tests passed locally and in Linux CI; macOS packaging passed. Numerical checks preserved tokenization on 144 inputs and sampled top-10 rankings for all eight queries. Tests cover repo isolation, shared relations, cross-type evidence IDs, lifecycle history, and upgrade failure readiness. The old backend and version helpers are removed; stale README assertions now match main.

The knowledge builder now uses gpt-6-luna with its existing xhigh reasoning and Fast service tier. The model switch passes all 27 focused configuration and Codex adapter tests.