feat: add bounded native Qwen3 CPU reranking

[forkwright]

Summary

• Add strict Qwen3 rank-profile admission and raw final-token [yes, no] execution over the shared causal body, with artifact-derived geometry and an independent F32/Q8 f64 oracle.
• Implement bounded native CPU reranking through the existing Reranker contract: artifact-owned template, trusted setup instruction, exact tokenizer compatibility, independent byte/token/batch limits, and one finite signed yes-minus-no score per input index.
• Extract the bounded immutable template renderer into templates for text and rerank, preserving text behavior and the default Stella/ModernBERT paths.
• Refuse configured tokenizer padding/truncation at all three native pipeline boundaries without changing ordinary tokenizer semantics. Extend the GPU-free debug/release graph witness.

Validation

PR and merged-main gates pass: 541 workspace tests (8 skips), 75 release kernel tests, and 207 minimal GPU-free CPU library tests in each of debug/release. Strict Clippy, formatting/structural checks, security, GPU-denied boundary, HIP code-object compilation/inspection and merged-main CodeQL pass. Independent synthetic body/head and end-to-end pipeline witnesses cover numerical sign/order, conditional instruction/role framing, distinct UTF-8-byte/token/batch limits, typed refusals and same-instance retry.

The merged tree is identical to reviewed head 4208bb777194b82875cba10917f6a5128dd30f8e. This is hosted combined proof; no local full-gate attestation is claimed.

Boundaries

No real-model payload runs, quality/throughput claims, GPU probes, deployments, service/mode/provider changes, or reindexing. No core trait or hipcore API/source changes. CPU and compilation evidence do not qualify hardware or resolve exact reranker conversion provenance.

Part of #150; does not close the foundation program.