The live per-frame scan visits only the set of locks with nonzero wait
counts, updated at each lock event: every wait-for edge originates at a
waiter, so uncontented locks cannot contribute anything. File loads replay
events without touching the set and are covered by the one-shot full-map
scan at end of load.
The scan consumes the final Waiting/SharedHolding flags of each lock
map: a thread blocked in a lock acquisition emits no Obtain and no other
event, so an unclosed wait at the last known state is a permanent wait
per the event contract. A failed try_lock never emits a wait, so it
cannot fabricate an edge. Locks recorded by clients with inverted
release ordering are skipped entirely, as their holder attribution is
ambiguous.
A thread that holds a shared lock and waits for exclusive cannot
progress while its own shared hold persists; this is reported as a
size-1 cycle.
Capture has serialized lock events since d6f32a083 and 3e3aa80fa, so
the out-of-order reconstruction machinery has nothing left to repair.
Mark now targets the recorded Wait/Obtain event index; the old reverse
scan could walk off the start of the timeline for a thread that never
waited.
