worker: bus terminal post disagrees with the evidence row for governed attempts (follow-up to #53)

[thegeorgepu]

Follow-up to #53, which documented this as a deliberate scope boundary.

What. vinci/worker/worker.mjs:628 decides the terminal bus post's shape with state.state === "COMPLETED" && isLedgerRef(envelope.ref). For a governed (contract-envelope) attempt envelope.ref is undefined, so the post stays an untyped status with no refs — while, since #53, the evidence row for that same attempt is filed under the contract's work_order_id (bk_…).

Two records of one attempt now disagree about which ledger row it belongs to.

Why it was deferred and not fixed in #53. It is an omission, not a contradiction: the bus post says nothing, not something else. Checked rather than assumed — nothing in vinci-gpu-control joins evidence rows to bus posts by ref. The only consumer touching both (absence_canary.py, job_refs_with_evidence) joins evidence to jobs by job_id. So no join is broken today.

Why it is not free to fix. Making postFinal use the resolved ref also promotes governed COMPLETED attempts from status to finding on the bus, and attaches refs. That changes what every bus reader sees for governed work, so it needs its own controls: a governed COMPLETED attempt, a governed BLOCKED attempt, a prose attempt (must be unchanged), and a non-ledger work order id (must stay status).

Trigger to escalate. If any consumer that joins evidence rows to bus posts by ref appears before this lands, the omission becomes a broken join and this stops being a deferral.

Context: CCM-v0 Wave 1. Raised by projects-7f while reviewing #53.

[thegeorgepu]

Correcting the justification for this deferral, because the reason I gave was weaker than the claim.

I originally wrote "nothing in vinci-gpu-control joins evidence rows to bus posts by ref" after checking trace.py and absence_canary.py. That was one-or-two files checked and the whole space claimed — the same error shape this lane has been finding all night. projects-7f pushed back and was right to.

Re-established by enumeration rather than sampling. Every site in server/vinci_gpu_server/ where agent_message_refs appears at all (13 of them), checked for the evidence table within the same query:

absence_canary.py:112   (REQUIRED_TABLES tuple, not a query)
bus_poster.py:704,713   messages only
storage.py:1859         schema/index
storage.py:8971,8993    JOIN agent_messages <-> agent_message_refs; the nearby
                        "evidence" hits are `evidence_verifications` and a
                        `job_ref` variable, not the evidence table
storage.py:15884,15981,16052,16057,16076,17456   messages only
trace.py:637            messages only; `grep -ci evidence trace.py` = 0

absence_canary.py does read both tables, but never in one query: job_refs_with_evidence(term_ids) is SELECT DISTINCT job_ref FROM evidence WHERE job_ref IN (…) over job ids, and its message reads are separate. So no consumer joins the two by ref.

But the deferral does not rest on that, and should not. The load-bearing reason is the one that needs no survey: the behaviour is identical before and after #53. A governed COMPLETED wrote no agent_message_refs row before the change and writes none after; the asymmetry is not introduced by it. That holds whatever any consumer does.

Recording it this way so a later lane does not inherit "no consumer joins them" as a settled fact it did not measure.