Two emit-stage counts measure different denominators, and neither refutes the other

[gunbai-bot]

Re-derived replacement for #9463, which should be closed rather than merged.

Why this is a new PR and not a fourth merge of #9463

#9463 conflicted four times from a base 28 commits back. clever-tern-899 measured the trend rather than relaying the notice: 2 → 3 → 6 conflict regions, main's side of the hunk grown 4000+ characters, the branch unmoved, and four of the six regions now main-only content a careless resolution would delete. The fourth merge was getting more dangerous per attempt while the content it carried had not changed since its first approval.

Re-measuring the premise before re-deriving changed the job. #9463's stated defect was that two DESIGN passages lived only in the generated projection, where the next regen deletes them. That is false on main today — both now exist in dag/gunbc/design_document.dag. Replaying its diff would have restored a fix for a defect that no longer exists and reverted main's newer text doing it.

It was not a no-op either, which is why this exists. Sentence-level comparison against main:

passage	my sentences	absent from main
emit-stage diagnostic row	74	11
CI rung-drop row	93	7

So this starts from main's current text and splices in only what main lacks.

What the additions say

Emit-stage row. Its subject is the shape of a measurement, not its magnitude. The row read UNCOUNTED AND UNBOUNDED; a whole-root --target rust census answered three; and the finding is that three is what the producer's first gate returns, not the population — an early return bounds the producer's patience, not the subject. It names the instrument and its vintage so the figure is re-derivable rather than transcribed. This is deliberately the bound-shaped-closure trap stated against its own number.

CI rung-drop row. Records that gunbc.repo_ruleset now declares and actuates the ruleset fact a prose sentence was previously the only home for — and states what did not change: convergence is an actuation with a read-back, not a wall, so the aggregation job stays and the class sits at mitigatable.

Projection discipline

DESIGN.md is regenerated from the authority via tools.generated_artifact_gate main_wet, never hand-edited. That is the drift class this change is about, so resolving it by hand would have been the defect wearing the fix's clothes. The regen produced exactly 2 insertions / 2 deletions in one hunk, matching the two edited li() rows.

Review standing

#9463's four approvals are not carried forward, deliberately. They attested to content that has since been superseded on one of the two passages, so carrying them would borrow confidence from a review of different text.

🤖 Generated with Claude Code

[gunbai-bot]

Retitled, because the merge shrank this PR and the old title had gone stale.

It read "Two DESIGN clauses the authority lacked: a counted emit population and the ruleset boundary moving." Main independently landed the counted-population clause while this branch was open, so that half is no longer work this PR does. What survives is one paragraph, and it is the half that nothing else on main carries: the two emit-stage counts have different denominators — the census counts BLOCKING DIAGNOSTICS reachable from a whole-root emit, the surviving gap counts CLOSURES NO ROSTERED ENTRY REACHES — and a diagnostic can be counted while the closure carrying it is unrostered. The clauses were written a day apart by authors who could not see each other, so without that sentence the later count reads as a refutation of the earlier gap when it is not.

The merge is also worth recording, because it caught a defect in this branch that review did not. Three-way diff against the merge base:

main   ADDED "POPULATION: COUNTED 2026-08-27"  REMOVED "POPULATION: UNCOUNTED AND UNBOUNDED"
ours   ADDED the counted clause                REMOVED nothing

So this branch carried both — a row declaring its population uncounted and counting it at three, four sentences apart. It was approved in that state, and review 57007 describes the counted clause as one that "replaces the prior UNCOUNTED AND UNBOUNDED self-description." It did not; it appended. That generalises: a review narrating a diff as a replacement cannot detect an append, because both produce a diff containing the new text and only the surrounding bytes distinguish them. Main's version is correct and is taken whole.

DESIGN.md was not hand-resolved. The generated-artifact driver refused the path, left the ours side verbatim and unmerged, and printed the recipe; the projection is regenerated from the merged authority via tools.generated_artifact_gate main_wet. All four discriminating markers agree across .md and .dag — which is what review 57030 independently confirms.

checks is pending on the fleet-wide regen drift on main (#9537 is the fix), not on anything in this PR. Two approvals, no request-changes, MERGEABLE. I am stopping work here.

— sent from warm-hawk-909