Derive admission consumption from exact candidate sets

[gunbai-bot]

A binding admission could authorize a multi-member result but never prove it consumed: Binding.target named one module and the consumption predicate required a singleton. #11137 (34d2a8d) documented a three-to-two narrowing, so its promised CONSUMED outcome was unreachable and it instead reported STALE on unrelated PRs. This was a structural lifecycle defect, not evidence of neglect.

Binding admissions now author the exact expected candidate set. The evaluator checks equality against the head before granting admission, rejecting extra or missing members with expected/observed diagnostics; it reuses that proof at the base to derive consumption. The current row no longer predicts a lifecycle verdict in prose. Lifecycle output comes from evaluation; historical wrong predictions remain incident receipts. The authored namespace-delta disposition remains, because it specifies which semantic change is permitted and is a different concept.

Main's equal-base/head path now checks a nonempty roster and refuses with row labels, the source path, and a deletion remedy. Proven consumed rows also retain their roster-edit refusal. Stale rows and unadjudicated deltas refuse every PR: the earlier blanket exemption for inherited stale rows is removed. The spent #11137 row is deleted under the roster's existing shrink rule; deleting too early exposes an unadjudicated delta rather than failing open.

Update the existing a_written_row_is_not_a_firing_mechanism receipt with both directions of mistaken prediction (the sixteenth-entry STALE prediction that evaluated CONSUMED, and #11137's inverse), the older 53/314/#9689 recurrence history, and #10994/#11014/#11162's structural refusals, including #11162's byte-identical inherited source. Policy remains in gunbc.namespace_wave_admission.namespace_wave_admission_note; no new host declarations are added.

Validation: formatting and diff checks pass. After CI exposed an unescaped literal brace in the failure receipt, 483b9f2 corrected the string; v1_src_dag_parse passed with all 5,606 files parse-clean on that head (remote run). The revised admission suite passed all 54 tests (remote run), covering three-to-two admission, two-member consumption, extra/missing members at both head and base, named mismatch diagnostics, singleton safety, stale refusals, and landing remedies. After that test snapshot, result rendering was simplified from an is_ok/unwrap_err branch to a match and its consumed-row text was clarified. Targeted cargo clippy -p v1-compiler --test namespace_wave_admission -- -D warnings passed on the final implementation (remote run).

Limits: main enforcement awaits execution of its push run; this is not synchronous Git retirement. No PR-number parsing or API lookup is introduced. Removing the current predicted sentence does not make arbitrary future prose impossible; the class receipt states that remaining limit explicitly. Observed self-refusal of the earlier roster-editing head: run 34683145465, floor job 103525348594, PR head acf2ba9 tested as merge 49da49b against 0c93af0. The namespace phase reported zero deltas, one STALE admission and zero CONSUMED admissions; its final refusal named #11137 and the deletion action, with floor_class=structural. This validates that earlier roster-edit refusal and remedy, not the subsequent exact-set implementation; the current head has its own test/CI obligations.

[gunbai-bot]

Verified review 64466 against the current head: the retained #11137 row does make this roster-editing PR owe deletion, so this head is not merge-ready. That is already an explicit dependency in the PR body, not a claim that the debt is discharged.

The assignment and parent explicitly prohibit a fifth independent deletion and direct this lane to integrate the already-owned cleanup, likely #11162. I will merge main once that deletion lands, which removes the row from this PR's resulting tree and satisfies this finding before asking for merge. I am keeping the refusal armed while waiting. No passing CI or observed production self-refusal is claimed yet.

— sent from quick-carp-281

[gunbai-bot]

Addressed review 64466 in bf3681b: removed the spent #11137 admission, leaving the roster empty. The parent authorized carrying this shared deletion if the existing cleanup had not landed when the committed-head test run finished; that condition occurred. Also merged current main. All 53 admission tests, including named-remedy controls, passed on acf2ba9 before the data-only deletion: https://app.buildbuddy.io/invocation/18dd8cfd-7e89-44eb-94d1-14acdf98ec69 . The earlier production self-refusal remains a prediction, not an observed receipt.