Implementation update — scheduler admission and durable dispatch
The first coherent MVP-06 slice is published in MR !11, branch ai/mvp-06-scheduler-quota, commit 25c5276.
It adds the one-subscription contention slot, tenant scheduler counters, atomic quota reservation/admission, deterministic payload-free dispatch publication, bounded 16-bucket quota expiry, and a real reconciler process. Public documentation and local/YDB integration coverage are included.
Local make ci, tagged compile gates, shell validation, Compose YAML parsing, and git diff --check pass. Real YDB Local, Compose, and queue behavior are intentionally left for the mirrored GitHub Actions jobs because Docker is unavailable on the development workstation.
This does not close #9. Remaining MVP-06 work: lease claim/recovery integration, cancellation, provider quota/entitlement observations and usage confidence, reservation commit/release at worker completion, and admin-safe aggregates.


Verification update — partition-contract CI fix
GitHub Actions run 30478731808 failed because the two new scheduler migrations requested a 16-partition floor while the architecture contract classifies those tables as tenant-first point tables with a one-partition floor.
Commit 68b7f53 aligns subscription_scheduler_slots and tenant_scheduler_counters with that contract. Load- and size-based YDB auto-partitioning remain enabled, so the tables can split as traffic grows without preallocating global-bucket capacity. A regression test now locks this distinction in place.
All local verification and both tagged integration compile gates pass. MVP-06 remains open until the mirrored GitHub workflow applies the corrected migrations and validates the resulting settings against a clean YDB Local instance.


Parent and architecture
Outcome
Provide deterministic admission, fairness, leasing and quota accounting for user-owned AI subscriptions without API-billing fallback.
Estimate
Scope
READY,PRESSURED,DRAINING,BLOCKED_UNTIL_RESETandREAUTH_REQUIREDstates.provider_reported,locally_measured,unknown.Policies
Acceptance criteria