A refusal that claimed admission passed when nothing evaluated it

[gunbai-bot]

harness_bind_candidate's bound-offer arm said "the serving offer bound and its
route is admissible". That branch never calls spark_first_party_admission, so
nothing evaluated provider-use admission for the turn: the sentence outran its
own execution, and a reader could take the producer EXISTING for the producer
having RUN. It did exactly that -- the sentence was relayed upward and reached
the operator as a statement about a decision nobody made.

The wording now carries the distinction the arm is about: the offer bound, and
admission cannot yet be evaluated because the turn's exact Work identity is
unestablished. It names the producer and says plainly that it has not run.

This is the correction from the operator-routed review of #11083, which merged
as bd32663. The route-and-policy slice it landed is not reopened: an
attributable allocation establishing a route's location, with observed and
operator-asserted kept apart, and a provider-use producer that reuses
product.fabric.provider_admission with DispositionNotEngaged instead of invented
clearances, both stand. What was wrong was a sentence about them.

First repair from the operator-routed review of #11083.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WT6FDJa7Nyh9SJv25sM1Ui

[gunbai-bot]

CI status, since the dashboard reports this as "2 failing" and asks for a fix: there is nothing in the diff to fix, and I have not pushed.

The two reds are one failure plus the aggregator. witnesses failed in 4s because it reads the lane results; the real one is required-witnesses-floor, and the runner itself classifies it:

required-ci: floor refused: TERMINAL-LEDGER REFUSAL cause=WireAuthorityUnresolved
  entry=src/v2/workflow/floor_terminal_ledger_wire.dag
  detail=MemoryStallRefusedPageThrash: this process refaulted its own pages
##[error]floor_class=infra signature=MemoryStallRefusedPageThrash exit=1;
  this is not a verdict about the diff

That is a phase refusal — standing=measurement_completed blockers=1 — on an entry unrelated to this change, which touches one refusal string and one annotation in harness_seat. The same signature has been refusing for several PRs across authors, and another lane established by replication on two hosts that it is not slot-specific. heal-generated-artifacts and required-witnesses-build both passed on this head.

I have re-run the failed jobs once rather than pushing. Pushing a no-op commit to trigger a fresh run would move the head, discard the approval on c54441d, and buy nothing — the refusal is not about the content.

If it refuses a second time I will stop and report rather than spend a third runner on it.

— sent from keen-owl-253

[gunbai-bot]

Second refusal, same signature, so I am stopping rather than re-running again.

required-witnesses-floor on c54441d: first attempt 40m44s, rerun 40m24s, both ending

floor refused: TERMINAL-LEDGER REFUSAL cause=WireAuthorityUnresolved
  entry=src/v2/workflow/floor_terminal_ledger_wire.dag
floor_class=infra signature=MemoryStallRefusedPageThrash
  "this is not a verdict about the diff"

heal-generated-artifacts (24m35s) and required-witnesses-build (28m19s) pass on this same head, both times. The second red is the aggregator reading the lane result, so this is one refusal reported as two.

What the floor is saying is that it could not measure, not that the diff is wrong — standing=measurement_completed blockers=1, class infra, on an entry this PR does not touch. The change here is one refusal string and one annotation in harness_seat.

I am not pushing a no-op commit to re-roll it: that moves the head, discards the approval on c54441d, and costs another 40-minute attempt against a memory stall that has nothing to do with the content. The same signature has now refused four times across two independent heads of mine on different runners, and was measured on main itself at a600fa07, so it is neither slot-specific nor this branch's.

Escalated to the lane that owns the runner memory envelope. This PR is content-complete and approved; it waits on that rather than on anything here.

— sent from keen-owl-253

[gunbai-bot]

RETROSPECTIVE RECORD — written after the merge, not a backdated authorization.

This PR landed without floor acceptance. Recording the facts so that anyone reconstructing what was integrated, or citing this merge later, has them:

• Candidate: head c54441d.
• Actual landing base: c586212 (The live probe already fetched the model path and threw it away #11115's commit), which is NOT the base SHA in this PR's metadata.
• Landing commit: 0c93af0, single parent c586212, merged 2026-09-12T06:36:33Z.
• Failed gate: required-witnesses-floor, and the aggregate witnesses, in run 34670582182 (attempt 3, event=pull_request). Floor completed FAILURE at 06:18:08Z, aggregate FAILURE at 06:18:13Z; build SUCCESS and heal SUCCESS on the same head. The merge followed the floor's failure by eighteen minutes, so this was a merge with missing acceptance rather than a failure discovered afterwards.
• All three floor attempts on this head refused identically: floor_class=infra, MemoryStallRefusedPageThrash, adjudication REFUSED standing=measurement_completed blockers=1.

WHAT THAT DOES AND DOES NOT ESTABLISH. The infrastructure classification explains why floor acceptance was unavailable; it does not supply that acceptance. The floor produced no verdict about this diff, so nothing was waived about its content — the required gate was unsatisfied, not satisfied-and-overridden. This merge is therefore not permission to land other infrastructure-red candidates, and not evidence that the memory envelope is repaired. Any subsequent exception is a separate operator decision, and the shape one would need is: the named candidate and landing base, the named missing gate and observed refusal, and a bounded scope.

No operator rationale or scope statement appears in the material inspected here. That is a statement about what this record contains, not a claim that none exists.

— sent from bright-eagle-728