The dedup now runs after the per-arm loop, so a missing reference path or a
missing field is reported as itself. A repeated reference arm is still
caught: the loop lets a well-formed reference through without erroring, and
the dedup reads the arms as written.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A field the format rendered and a calc read was drawn twice, so the operand
shown could differ from the operand computed — the same disagreement the
dotted-path rule already fences, in its plainest spelling. The README
carried it as a caveat, which is what a rule like this exists to remove.
A calc operand now joins the names an expansion holds, so it is drawn once
and every later read of it — the calc, and any bare token spelling it —
sees that draw. The hold stays per expansion: each repeat iteration and each
nested template draws its own.
expand reads a calc's operands before the call and hands over their values,
so a builtin takes (emitted, operands) rather than the sibling fields, and
the evaluator indexes that slice instead of walking the node tree. That is
what keeps the draws a local. Handing them to a builtin instead lets them
escape through an indirect call, which put two maps on the heap for every
held format, calc or not: BenchmarkBound went 56 B/5 allocs -> 744 B/10.
Measured against main: Bound 359 -> 349 ns at 56 B/5 allocs, unchanged;
Calc 416 -> 577 ns and one more alloc, which is what the correlation costs.
Validation is unchanged: t.bound still carries only the levels a dotted
token reads, so the overlap fences fence exactly what they did.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An alternation picks its arms evenly, so {a|a|b} was a covert 3:1 skew — a
third spelling of what a choice's weight is for, reachable only by
experiment. The check runs over the arms as written, before a reference arm
is let through, so {..a|..a} is caught too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A builtin could be spelled as prep (args parsed once, at compile) or call
(args parsed per render), reaching the same closure by two routes. Nothing
enforced the "exactly one of" invariant: supplying both silently dropped
call, supplying neither compiled and then panicked inside a render, and a
test existed only to police it.
prep does everything call did, so it is now the one spelling. The zero-arg
shapes the README already names — a derivation reads the output emitted so
far, a generator reads only the rng — lift through derive and generate, so
the registry reads as that taxonomy. iban stops looking its country up per
render.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
M
Mikael Göransson