[Denny-Lin]

An unexpected runnable-task stall could report watchdog_observed: false because that legacy field only recognizes expected watchdog tests. Preserve existing exit statuses and fallback decisions; add watchdog_expected and nullable watchdog_triggered so telemetry distinguishes normal, expected and unexpected exits. No BPF program, model, slice or default changes.

Validation: Linux ARM64 compilation and three real QEMU exit cases passed. The intentional unexpected-stall loader returns status 1 as required. A fresh 250-second, eight-vCPU Lerwatt integration passed across all three clusters, including default standalone runs before and after, consistent shared maps, all 30 actuator values restored, no residual journal, zero throttle steps and zero bad writes. Source and binary hashes plus compact results are in results/watchdog-status.

The first integration collector failed on bpftool's raw byte-array key. That attempt remains failed, with its raw dump and cleanup evidence retained. Only semantic BTF decoding in the test tool changed; the complete fresh rerun passed with identical product binaries. Full raw records are archived locally with a published SHA-256 manifest.

This improves diagnostics, not the unresolved natural watchdog stall reported in the model-confirmation results. The unavailable-enum compatibility path was not exercised on this kernel. This single simulated-board integration establishes activation and cleanup, not prediction, throughput or energy superiority. Defaults remain unchanged.