fix(pflash): score the user query and reject non-finite scores

[Graffioh]

Summary

• maps the final user message into the drafter-rendered token stream and uses its last eight tokens as the scorer query
• fails closed if that user query cannot be mapped, rather than falling back to the assistant/chat suffix
• keeps scorer-query selection separate from DFLASH_COMPRESS_QUERY_TOKENS, which remains the lexical-anchor window
• propagates the query window through local Qwen3/Qwen3.5, layer-split, and IPC paths
• assembles scorer-Q windows that cross a 4,096-token forward chunk boundary
• checks graph execution and score readback, rejecting NaN/Inf values before top-k reduction

Why

The previous fixed lookahead=8 window landed on the rendered assistant suffix rather than on the question. On Strix Halo/gfx1151, corrupted HIP forward values could also reach the scorer; NaN comparisons then left earlier maxima in place and silently produced a selection.

Validation

• rebased onto Luce-Org/lucebox@ef82fadf
• HIP Release build for gfx1151: dflash_server, test_tokenizer_harness, and test_server_unit
• HIP_VISIBLE_DEVICES=1 ./build-hip-gfx1151/test_server_unit: 439 passed, 0 failed
• live 16K Strix Halo smoke:
  • uncompressed control: 2/2 correct
  • PFlash: fails closed with explicit non-finite score diagnostics
  • query mapping log confirms the actual user-query window, e.g. [16374, 16382), rather than the rendered assistant suffix

The underlying gfx1151 numeric corruption is intentionally not hidden by this PR; this makes it observable and prevents quality results from being reported from invalid scores.

Paired long-context benchmark campaign: Luce-Org/luce_box#85.