Publish training-free choice-token readout matrix

[weitianxin]

Summary

• expose training-free choice-token readout as --readout choice, with legacy aliases retained
• evaluate frozen 4B/27B, Pointer 4B/27B, and Direct-Token 4B on one matched matrix
• select stack weight and temperature only on Transfer-v9 development, then freeze them for held-out Transfer, Typed Decisions, and JevJudge text
• publish validated JSON, a README SVG, complete method tables, and a deterministic 23-page merged technical report
• support all 52 one-token option IDs and memory-linear attention for long JevJudge text records

Headline results

• Direct-Token 4B tuned stack: 79.83% Transfer test, 67.65% Typed, 59.25% JevJudge text
• Pointer 27B tuned stack: 89.10% Transfer test, 73.30% Typed, 90.04% JevBench; native remains best on JevJudge text at 66.44%
• native parity gates reproduce the canonical README rows before held-out evaluation

Verification

• 97 relevant API/readout/evaluator/builder/playground tests passed; external-server tests require a running endpoint
• 5 report producer/consumer, in-place rebuild, page-preservation, and byte-determinism tests passed
• package wheel and sdist build succeeded
• result JSON/SVG and the 23-page PDF rebuild byte-identically
• full-suite run reached 455 passed; remaining failures are from the cluster eval environment loading CUDA-only causal-conv kernels in CPU tests or requiring an external HTTP server