Add training-free option-letter readout

[weitianxin]

What changed

• add Cygnet-compatible exact A-Z next-token readout for frozen bases and JevAny adapters
• expose --readout letter in eval, decide, serve, the Python loader, and latency tooling
• support an optional fixed log-linear blend with the native pointer readout
• add prompt-control escaping, context checks, selective untied-head row loading, provenance, tests, docs, attribution, and machine-readable results
• preserve the existing external accuracy chart, including TypeSafe Jev 1.13 at 72.7% Typed Decisions accuracy and JevAny-Qwen3.8-27B at 72.8%

Measured result

Same 231-question JevBench public diagnostic, T=1:

• 4B native 80.09%; letter 81.39%; fixed 50/50 blend 81.82%
• 27B native 89.61%; letter 89.61%; fixed 50/50 blend 90.04%

Transfer-v9, 1,046 clean knowable decisions:

• 4B native 79.16%; letter 75.72%; blend 79.06%
• 27B native 86.23%; letter 84.23%; blend 86.90%

No positive delta is statistically significant. The result supports letter readout as a paired domain ablation, not a universal replacement. On one warmed H200 diagnostic, letter-only median latency was 1.05x/1.07x faster at 4B/27B; blending nearly doubled median latency because it runs both prefills.

Cygnet's official 73.70 is documented separately as a four-axis composite over 1,624 open and sealed decisions, not accuracy.

Validation

• 92 targeted tests passed
• real JevAny-Qwen3.5-4B checkpoint smoke passed on H200, including hostile chat-control text
• site validation passed: 27 pages, 1,095 links, 239 assets
• JSON and SVG parse checks passed
• git diff --check passed
• independent final review found no blocker