[tannevaled]

#874 gave cmd/specdelta a refusal for a run whose measured set does not cover
the surface a declared version move touches. The refusal was right; the exit it
advised was not reachable.

The advice could not be followed

-dir took one directory, and the refusal said "Widen -dir". For the
move this repository is actually making, the census says where the surface is:

OPENED  language/variables_spec.rb:366  >=1 examples  ruby_version_is "4.0" do
OPENED  core/binding/implicit_parameters_spec.rb:3  >=6 examples  ruby_version_is "4.0" do

3.4.1 → 4.0.9 spans core/ and language/. Every one of the 50 opened
files is under one of the two, and none is outside them — so the surface has an
exact, small description. But from -dir core the only widening available was
the whole corpus: library/, optional/, command_line/, security/, and
hours of sweep for files the move never touches.

That left -partial as the only reachable exit, which means the headline figure
for #799 would have been published stamped PARTIAL — a derogation used not
because the surface was really incomplete, but because the flag could not
express a complete one. A derogation that is the normal road is not a
derogation, and a gate whose honest exit nobody can reach is not a gate.

-dir names a set, and the refusal names the right one

$ specdelta -root ~/ruby-spec-corpus -move 3.4.1:4.0.9 -dir core ./a ./b
specdelta: the move 3.4.1 → 4.0.9 touches 1 file(s) that -dir core does not measure:
    language/variables_spec.rb
  A number from this run would not be a number about the move. Either
    -dir core,language
  which measures the whole surface, or -partial to say so on purpose —
  the report is then stamped PARTIAL.

The suggestion is computed from the census, top-level suite per uncovered file:
50 exact directories would be unreadable and would go stale the moment the
corpus moves, while core,language is both the honest surface and something a
reader can retype. It reaches core/ and language/ and not library/,
which the test asserts — a suggestion that over-reaches would re-create the
problem it is fixing.

Two ways a set can be read wrong, both tested

case	why it matters
-dir core,core/set	the same file reached twice would be counted twice, on both sides: the totals would then move with how the caller spelled -dir rather than with the engines
-dir core,	an empty element must not be read as the root — a trailing comma would quietly sweep the whole corpus

And the control beside the refusal: retyping exactly what it advised is
accepted, exits 0, and is not stamped PARTIAL.

Refs #799

🤖 Generated with Claude Code