Docs: settle the vera judge command plan

[luca-belli]

Summary

Settles the decisions needed before vera judge can be built, so that
implementation is a mechanical mirror of #190 rather than a design exercise.

Docs only — no code. Stacked on #191 because the legacy-scripts paragraph this
extends was introduced there.

Decisions recorded

Evaluation output is named for the target, not the rubric —
evaluations/<target>/, matching c_<chatbot>. Only the path segment changes:
judging.rubrics keeps its name under AD-20, which exists to prevent exactly
this kind of schema break, and --rubric already takes a target name. Phase 4
notes that the segment names the rubric-providing target, which --rubric
allows to differ from the run's own target.

judge.py is carved out of Phase 1's script deletion. vera judge ships
without --resume because the resume contract is deferred, so judge.py is
retained solely as the resume entry point until vera resume exists, then
deleted. Scoped narrowly on purpose: a legacy script kept alive for one reason
otherwise becomes permanent.

--conversations takes exactly one folder, preserving today's
judge.py --folder behavior, while keeping the flag and config field
list-shaped per AD-20's reasoning. Judging several folders is already
expressible — judge each, then combine with vera pool.

vera judge --target all is deferred to Phase 4, with a Done-when bar
rather than a floating note. The resolution is clean today (N separate length-1
runs); the blocker is that until the <target>/ segment exists, those runs are
distinguishable only by timestamp, so output cannot be attributed to a rubric.
--target all keeps its full meaning for generate.

Phase 1 judging keeps the existing <gen_run>/evaluations/j_* layout,
renamed in Phase 3.

Not changed, deliberately

• judging.rubrics — renaming it is the schema break AD-20 prevents
• --rubric — already a target-name selector, symmetric with --personas
• The spine's adopted rules, beyond the same path-segment rename for consistency

Consequences for the implementation

--target XOR --rubric for judge mirrors --target XOR --personas for
generate exactly: same exclusivity rule, same all semantics, different field
sets. ResolvedTarget already validates and resolves all five fields, so judge
needs no new manifest code — and the resolution helper should be parameterized
by which fields the target supplies rather than copied per command.

With single-conversation mode dropped, resume excluded, one output rule, and
set_debug moving to the CLI, the judge domain-prep step is no larger than #190.

🤖 Generated with Claude Code