speculative: add opt-in prompt lookup for DFlash2

[bri-prism]

Overview

Add opt-in prompt lookup to DFlash2 with LLAMA_DFLASH2_LOOKUP=1. When the committed output matches a prompt span, use its continuation as the draft and skip the neural noise-block forward. The target still verifies every proposed token. Misses, ambiguous continuations and changed prompt prefixes fall back to neural drafting. Disabled by default; target/drafter weight sharing is unchanged.

The first round requires a unique two-token match. Later rounds use the longest matching suffix of 16-64 tokens, bounded by the original prompt and requested draft depth.

Validation

• Release llama-server build and standalone C++ policy checks passed: unique/ambiguous matches, depth boundaries, prefix mismatch, suffix matches and identical/conflicting ties.
• M5 Pro, one greedy request, 512 prompt tokens, 129 emitted IDs (128 decode tokens), warmup plus five measured repetitions per arm: median decode 75.72 -> 154.29 tokens/s (2.04x). Median prefill plus decode 3.056 -> 2.186 seconds.
• All output IDs in every measured run and warmup matched the serial control. Both arms used the same build and retained the same idle/thermal gates.
• This is a copy-heavy coding prompt, not evidence of general-workload speedup. Multi-request, long-context and mixed-hit stress validation remain pending, so the feature stays opt-in. Timing is paused; no additional performance claims are included.

Requirements

• Contributing guidelines reviewed; this submission targets the maintained fork.
• AI usage disclosure: YES. Codex implemented the change and assisted with validation and this description.