The Whole-Unit Read Is Scoped to Carriage, and the Evidence Says It Should Be Scoped to Text Shape #1437

Evidence from PR #1435 that the whole-unit read is scoped to the wrong property, and a measurement that
separates the two candidate causes.

What happened

GOVERNANCE.md "Communicating with the User" is carried, so every change to it owed a whole-unit
adversarial read. It got eighteen. repo-config/README.md is not carried, so it owed none, and it got
diff-scoped reviews instead.

The Labels section of that uncarried file was then broken four times in a row, each time by a fix
to the previous break:

1. decision moved out of the four mutually exclusive triage classes and made a marker. The invariant
   sentence seven lines above still said an issue carries exactly one of them.
2. The marker paragraph said it "sits beside one of the four", making a class mandatory, which left the
   pure parked question GOVERNANCE.md names with no legal label set.
3. The invariant gained the exemption. The marker paragraph had by then introduced "class" as the noun
   for the four, colliding with the sentence one blank line below it, which uses "class" for the
   review-finding labels introduced and pre-existing.
4. Two more in the same paragraph: a pronoun binding that made an issue mark every question rather than
   the marker marking one, and "has a session present at handoff" reading "present" as an adjective,
   which collides with that rule's own presence vocabulary and turns an obligation into a precondition.

Each fix was verified. Each verification was scoped to the sentence that changed. The first read of the
section as one whole text found three of the four, immediately.

The measurement

Across this branch, the surfaces that carried the same rule:

The generated copies never drifted, which is #1317's argument. The hand-written surfaces that owed a
whole-unit read had their defects caught in-branch. The hand-written surfaces that owed none drifted
seven times between them and were caught only by a reviewer or by a whole-tree diff pass, which is a
weaker net and a later one.

What this suggests

The rule in GOVERNANCE.md "Verification Discipline" scopes the whole-unit read to carried content,
and the reason it gives is that a carrier reads the unit whole and cannot fix what it finds. That reason
is real. But the evidence here is that the whole-unit read is not actually a remedy for carriage. It is a
remedy for a multi-sentence text edited one sentence at a time, which is a property of the text and
not of who reads it. Carriage is one situation where that hurts. An eight-sentence section in an uncarried
file is another, and it drifted more, not less.

Two things worth deciding:

1. Whether the whole-unit read's trigger should be the shape of the text rather than its carriage, for
   example any section a change edits that exceeds some size, whatever file it sits in.
2. Whether canonical_review.py should be able to record a pass over a unit in an uncarried file, since
   today the ledger's unit set comes from the carry manifest and there is nowhere to record such a pass
   even if one is run.

Filed rather than acted on, since changing what the gate covers is a fleet-wide decision. The immediate
cost is bounded: the four breaks above were all caught before merge.