-- `?h cond a b` — cond MUST type-check to `b`. The verifier enforces
-- this (ILO-T038) so non-bool conds can't silently fall through to
-- the then-branch.
--
-- Background: 0.12.0 had a silent-truthy bug where a parenthesised
-- prefix-comparison like `(> p 0.5)` in cond position was mis-parsed
-- as a zero-param inline lambda (`(>n;body)` shares its leading
-- tokens), the synthesised fn-ref reached `?h` as a non-bool, and
-- the runtime always took the then-branch — ml-tabular rerun11's
-- logistic-regression classifier silently scored 25.75% instead of
-- 84.6%. 0.12.1 fixes the parser to require a `;` body separator
-- before treating `(> ...)` as a lambda, and adds the verifier
-- check below as defence-in-depth.
-- Canonical form 1: comparison-derived bool bound to a name first.
-- The cheapest token-count shape for non-trivial cond.
clf1 p:n>n;c=> p 0.5;?h c 1 0
-- Canonical form 2: paren-grouped prefix-comparison inline. After
-- the parser fix this types correctly as `b` and dispatches right.
clf2 p:n>n;?h (> p 0.5) 1 0
-- Canonical form 3: brace-delimited ternary. Grammar-delimited arms
-- make the structure impossible to mis-parse, at a small token cost.
clf3 p:n>n;c=> p 0.5;?c{1}{0}
-- run: clf1 0.7
-- out: 1
-- run: clf1 0.3
-- out: 0
-- run: clf2 0.7
-- out: 1
-- run: clf2 0.3
-- out: 0
-- run: clf2 0.5
-- out: 0
-- run: clf3 0.7
-- out: 1
-- run: clf3 0.3
-- out: 0