v0.12.2

Kubernetes ResourceQuota pressure now waits with shared exponential backoff instead of failing successive worker intents. Workers retain the same intent across capacity deadlines, remain cancellable, and wait for the previous reservation to be cleaned up before retrying. Permission failures, image failures, and unknown execution outcomes are not replayed as capacity errors.

Runner diagnostics and the UI now distinguish logical slot reservations, execution-pool Pods, and namespace quota snapshots. Preparation stages, waiting duration, create attempts, rejection counts, and recovery are visible without inspecting task content. Independent heartbeat reporting preserves newer observations when long polls arrive late, with negotiated compatibility for upgrades and rollbacks.

Validation: full PostgreSQL/backend regression, race checks, static analysis, frontend build and desktop/mobile browser smoke. An isolated K3s regression verified actual quota rejection, quota increase, concurrent cancellation, Pod deletion and quota reclamation, and zero remaining test Pods. No task content was used in validation.