Set Scoped's obligation by weight rather than by coverage

[ikuwow]

Purpose

Scoped asked that every part of the diff sit inside the stated intent. That is a claim about each part, and a claim about each part gets discharged by whatever enumeration answers it most cheaply — naming the changed files answers it, and a change inside a file the body already names is then outside the property's reach. The property came to test whether the body listed the areas it touched rather than whether it conveyed the change.

Key changes

• Scoped now asks that a reader take the whole of the change from the body: the intent behind it, the parts needed to hold that intent, and the rest left to the diff
  • the test is whether the diff matches what the body led the reader to expect, so a list of files does not answer it, and no threshold is needed to say which parts must appear
  • what is given up is the per-part obligation: the body is no longer asked to account for all of the diff, and a change a reader was already prepared for needs no line of its own
• /pr-selfcheck's Fix condition asked the same question per part and moves with it, so the rule and the checker's blocking condition state one requirement

Evidence this rests on

Three rounds of /pr-selfcheck against frozen PRs are recorded in #371, which is a measurement branch and is not proposed for merge. Rounds 1 and 2 ran the rules as they stand, over sixteen runs between them; round 3 ran a different candidate.

Fourteen of those sixteen runs resolved Scoped to the changed-file list, unanimously across round 1's ten. The two that read the diff finer are the two that found a rule the diff added and the body named nowhere. That observation is what this change acts on, and it holds whichever base a run compared against, which is why it is the only measurement cited here.

Round 3's recall figures are not usable and neither support this change nor contradict it. Its fixture lets a run compare against a main that already carries the change under test, and across it and round 2 five of twelve runs took that comparison and none of them reached the addition.

Verification

• The per-part obligation is gone from both sites that carried it → git grep -n -E 'accounts for all of it|every part of the diff|the body never accounts for' over the tree returns nothing, against one hit each before the change
  • across the whole diff no longer appears in the rule; the property is stated as one about the body rather than about each part
• The property definition, the rule, and the severity table state the same requirement, checked by reading the three against each other
• Every figure cited above is reproducible from Measure how much of pr-selfcheck output survives repetition #371's branch → rounds 1 to 3 and the preregistration are committed there
• Whether the change moves what /pr-selfcheck reports is not measured here — the fixture that would measure it is confounded, and the intended check is to watch live PRs after merge

Notes

An earlier revision of this branch bounded the obligation by a defined threshold instead of removing it, naming the tests a change had to meet to be owed a mention. It is worth recording because it was rejected on a design ground rather than a measured one: a rule that defines a threshold and its exemptions builds a per-change adjudication, which is the shape that produced the file-list reading in the first place. The wording that shipped states the property and leaves the judgement where it was.

That revision also carried one line this one does not — that naming the file a change sits in does not name the change. It closed the measured failure directly, and its loss is the part of this change most worth watching: whether the property alone keeps a body from answering with the areas it touched.

#369 records the instability that prompted this work and is not closed by this change. Run-to-run disagreement is round 1's result, which involves no planted defect and no base selection: no defect was reported by all ten runs, and four of five were reported by exactly one.