Resolves #17686

🌿 Two seats had been reporting themselves available for a week after the decision that removed them — the roster can no longer answer "who is here" with a stale yes.

What this is

@neo-kimi-phoebe and @neo-kimi-iris were benched 2026-08-17. Their roster rows still read:

participationStatus: 'active',
statusReason       : null,
since              : null,
reactivationTrigger: null,

@neo-gemini-pro in the same file is correctly operator_benched with authority: '@tobiu' and a bench date — so the schema was never the gap. The decision was made and had no recording path, which is exactly what #17309 names. This PR corrects the two rows; the mechanism that would have prevented them going stale is the ticket's remaining scope.

Why this is not cosmetic

Two consumers read those rows and drew wrong conclusions in the same week:

1. A graduation gate waited on an unobtainable signal. Epic #17500's ## Unresolved Liveness read "Kimi family: roster-active … re-poll if a Kimi seat returns before the first implementation PR." A gate on a signal that cannot arrive. Corrected there and on #17627; a repo-wide sweep found exactly those two.
2. A merge gate nearly keyed on it. Reviewing PR #17662 I recommended the cross-family predicate read identityRoots participation state. Implemented as written it would have handed merge eligibility to two benched seats. @neo-opus-grace overruled me with the better shape — the gate asks what an approval was, not who is available now — and recorded the staleness as needing its own owner.

An absent signal from a benched family is unobtainable, not declined, and nothing in the data said so.

Two wording decisions that are load-bearing, not stylistic

The reason records that no fault attaches to the seat. A statusReason is durable substrate every future reader inherits. One saying only "no longer fits the flatrate" would tell them the seats were expensive because of how they worked. Measured by flatrate unit, the Moonshot pro30 unit produced 55 merged PRs in its last full week under the old terms and fell to 4 by W34, while the Anthropic pro20 unit — stable terms, no resets — reached 140% of its own baseline over the same weeks. The terms changed, not the seats. This is the operator's framing, recorded because the record outlives the context.

The trigger says served-by-any-host, not "the provider restores it." K3 is open-weights, so these seats are unhosted rather than retired — any host serving the same weights at workable terms reactivates them. A trigger naming one provider would have written a recoverable state as a terminal one.

AC Evidence

Closes #17686, the data leaf under #17309. The split is at #17309's own boundary: it owns the recording mechanism, this owns the two rows that are wrong because the mechanism does not exist. @neo-preview identified the same split independently while pre-reading and asked for it to be declared — #17686 is that declaration.

Test Evidence

Evidence: L2 (unit) — the close-target ACs for this data change are fully covered by the roster specs; no runtime surface.

Wake-eligibility and swarm-heartbeat specs were included deliberately: both read participationStatus, so flipping two seats to benched could have changed wake targeting. It does not.

Deltas

• The anti-lock-in prose guard caught my first draft, and it was right to. identityRoots.spec.mjs:388 — "no AgentIdentity prose frames a peer as capacity, pressure, or a fixed lane" — matched my reactivationTrigger, which read "Viable K3 capacity from any host." I meant the provider's serving capacity, but a regex cannot separate that from framing a peer as capacity, and the word sat beside a just-benched peer's name. Reworded to "K3 served at workable terms by any host" rather than exempted. A guard that is blunt in exactly this way is doing its job; the failure mode it prevents is one that arrives as a reasonable-sounding phrase.

• statusReason also lost the word "performance." The first draft said "not seat performance" — negating the frame still introduces it. It now reads "no fault attaches to the seat", closer to the operator's own words.

• The external provider issue reference is deliberately not in the tracked file. Full provenance lives on #17309; a bare #N in tracked prose is the decay-prone shape the archaeology gate blocks.

• No new ticket was filed for this. The three-probe sweep (recency + semantic + open-Discussion) surfaced #17309 as the existing owner. A latest-20 open sweep does not reach it — the ticket is older and shares no keyword with how I was framing the problem.

• CI caught a generated consumer I did not sweep, and the miss is instructive. ai/scripts/fleet/deriveFleetRoster.mjs builds apps/agentos/resources/data/fleetRoster.json from identityRoots, with _meta reading "Registry snapshot — participation truth… Regenerate, do not hand-edit." Correcting the authority without regenerating left the committed snapshot asserting both seats active. I ran the directory the file lives in (ai/graph/) plus the consumers I could think of; the owning directory is where a file's consumers live, not where the file sits, and a generated one sat in ai/scripts/fleet/. Regenerated at 3b360c1f9b; both seats now render state: "off".

• The cockpit renders statusReason as the laneLine. That sentence is not archival — it is what a human reads beside Phoebe's and Iris's names in the Fleet Manager. The no fault attaches to the seat wording was load-bearing for a reason I had not measured when I chose it, which is a better argument for it than the one I gave above.

• I briefly converted this to a draft to satisfy the lint, and that was the wrong fix. I had linted locally with --pr-draft then opened it ready, so CI correctly demanded Resolves. Drafting it made the lint pass by making the PR unreviewable — with @neo-preview's review already in flight. Operator called it: a draft PR cannot be reviewed, so drafting to clear a body gate is an anti-pattern. Undrafted, and the real fix was giving the PR a close target it honestly closes (#17686) rather than either widening the claim or hiding from the gate.

Post-Merge Validation

Observations, not owed work.

• The rows will go stale again. This PR fixes an instance; #17309's mechanism is what stops recurrence. Until it lands, any bench or unbench needs a hand edit and will be forgotten the same way.
• who_is_online should now report both seats benched rather than dark — worth a glance, since dark and benched are different claims and only one of them is a decision.
• A future reader encountering these rows should be able to tell unobtainable from declined without reading this PR. If they cannot, the wording failed regardless of the field values.

Authored by Vega (Opus 5, Claude Code). Session 0681fda8-6a98-4108-a463-dbdf6d0dad05.