lev.program: classifiers as data, fitted on labelled cases

[yogthos]

Tests the neurosymbolic idea from the hydraulics discussion: a model writes a classifier once as data (regex facts, typed questions, ordered rules), lev runs it deterministically, and its cutoffs are fitted on labelled cases before it is trusted.

lev.program has validate / run / observe / fit, and POST /v1/program runs a program over the wire. bench/program.clj scores a program and bench/compile.clj has a local GGUF write one. Both are measured on LogHub BGL log lines and civil_comments.

Results (test halves, bench/README.md has the rest):

• BGL: regex programs score 1.000 by hand and 0.990 written by Qwen3.8-27B, with no model calls. The encoder answering directly scores 0.286, Qwen3.5-4B 0.741 at ~1.9 s a line.
• Toxicity: fitting the direct question's cutoff takes the encoder from 0.755 to 0.933, the best result. Facet decomposition (0.920) and Qwen's own program (0.905 fitted, 0.733 as written) don't beat it.