recurring_failure_mode: a check and its subject removed in one edit, so nothing remains to object

[gunbai-bot]

One new recurring_failure_mode row. No code, no other files, no roster edit — membership in this ledger is the directory, and the receipts are consumed by the existing authored fold.

The class

A check and the thing it checked are removed in one edit, so nothing remains to object. A change removes a field, a declaration or an arm, and the same change removes the assertion that read it — because that assertion no longer compiles without it. No artifact survives that is false; no execution survives that can be inspected. The removal presents as the removal of the subject, which is what appears in the diff and what the change is described as. The wall leaves with it and appears nowhere as a loss.

Why it survives review, and it is not inattention. The question a reviewer asks about a removed field is whether the fact is still carried somewhere. That is the right question, and in the specimen it was asked and correctly answered — the grounding survived structurally, as an input to a content hash. The question nobody asks is whether anything was asserting the field. A field can be redundant as data and load-bearing as a subject, and the two are answered by different questions.

Recognition rule, diff-grain: for any change removing a field, declaration or arm, read the deleted lines of the same change for a predicate over the removed name. If they contain one, a wall left with it, and the change is not complete until that claim is re-made at whatever carrier the fact moved to.

Specimen

gunbc-private#102 with public #11194. PermittedJurisdictionsEnumerated landed without a basis field a private consumer constructed it with; population_is_the_clearance_roster ended with a declaration_ref_eq over that basis, so dropping the field dropped a check while "dropping basis loses nothing" was the stated justification. Three readers asked the fact question; none asked the subject question. The loss was found while re-making the claim elsewhere, not while reviewing the removal.

The claim was re-made wider than the one deleted: a_ruling_change_changes_the_revision tests the binding rather than a field's presence and, with the existing clearance conjunct, covers both inputs to posture_revision_hash where the deleted assertion covered one. Its red is measured, not argued.

Rung, ceiling, trigger

• Rung 1, mitigatable, honestly: nothing in the corpus detects it; it is silent wherever nobody re-reads.
• Ceiling 2, derived: whether a change's deleted lines assert over a name that the same change removes is decidable from the diff alone — both halves are in one artifact. It does not reach 3 or 4 because the substrate does not tie an assertion to its subject, so the state stays writable and the wall is a census rather than a construction.
• Trigger, named as the capability: a diff-grain census enumerating, for every removed declaration/field/arm, the removed expressions that referenced it, and refusing a change where such an assertion disappears without a replacement naming the carrier the fact moved to. Hand-authoring one census for one subject discharges nothing (§4b(3)).

Bounded against its neighbours, by reading their invalid states

Filing into a row whose invalid state a receipt does not exhibit is what #11245 was refused for earlier today, so these were read rather than recalled:

row	why this is not it
executed_conjunct_discriminates_nothing	its conjunct exists, executes, is gated on — "deletes cleanly: remove it and every witness stays green". There is an execution to inspect, a mutation that exposes it, a fixture that retires it. Here there is no conjunct: nothing to mutate, and no fixture retires what does not exist.
dissolution_trigger_cites_a_mechanism_that_is_later_deleted	the exact inverse — the citation survives its referent and becomes a false artifact still reading as tracked. Here both die together, nothing false is left, which is why the rule must read the deleted lines.
a_positive_control_certifies_the_defect_it_exists_to_catch	a control present and inverted; this one is absent.
unbacked_execution_claim	a claim whose referent never existed; this is an assertion that existed, was true, and was removed with what it read.

No sibling sweep is claimed — one specimen, and the row says nothing about the population of such removals in the corpus.

Verification

gunbc compile over the row's closure: rc=0, compiled: 1 files emitted, 96 advisory diagnostics, 0 blocking — on a binary built from this tree and proven the right vintage first by compiling a public file that a pre-#10850 binary refuses (rc=0, zero unparseable refusals). That control exists because staleness bit in both directions today: a stale binary refusing 82 current public files, and a current binary rejecting an older tree's func form. Both surface as rc=1 and neither names which side is stale.

Checked before filing, since a near-duplicate face is what this ledger exists to avoid: no row in main carries this class, and of the 20 new rows across 80 open PRs the two whose names mention a check are an unreached entry body never typechecked and a precondition check downstream of its effect — a check that exists and is misplaced, not one that is gone.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WqUFCjC1JF3yREmbwUTJKh