code-v48

Publish the real three-seed author/beam ablations and independent rank/work measurement, with raw rows, paired summaries, source hashes and logs. Fix raw fixture references to use portable repository-relative paths.

just test-fast: 259 tests OK in 227.655s, no exclusions. The complete affected slow lane is running and retry follows sequentially; neither is counted as passed here.

Reachability-trained full scoring accepts 9/12 shadow author and 10/12 shadow beam cases, versus 0/12 and 3/12 without neural scores. Novel acceptance is 0/9 and 1/9. Novel beam regresses from 6/9 with canonical labels to 1/9 with reachability labels; both separately fitted pipelines are published. Five of six fits selected zero exact/gap weights, so most named bonus ablations are null contrasts.

The separate 180-row rank measurement reproduces strong shadow ordering, recipe-dependent transfer, and greater eager guided signature work than streaming enumeration. It does not establish a guided signature-work advantage.

This supplies #70's requested paired results and documents classical candidate/signature computation. All six requested issues now have implementation or measurement evidence; integration slow/retry validation remains pending at this prerelease.

Report, raw data, exact validation evidence and limitations.