Simulation rewind/edit loops: test that the abandoned future never leaks into the waveform

When a tool rewinds a running simulation to a checkpoint and re-runs it (e.g. to try an edit), the waveform must contain the prefix up to the checkpoint plus the new run, and nothing from the run that was abandoned. Two things to check. (1) Open files across fork-based checkpoints: if the checkpoint is a forked child, it shares file descriptors and offsets with the parent unless it reopens them, so a resumed child can interleave writes with the parent's dump. Have the child reopen its output files privately, and verify that a restored run rewrites the files byte-identically. (2) Add a negative control to the exactness test: deliberately let the abandoned branch's writes into the reference, and make sure the comparison then FAILS. Without it, a comparator that only checks the common prefix passes either way. Also expect checkpoint-join in batch simulators to discard checkpoints after the join point by default (keepfuture off). If users want to compare the old and new futures, keep the old tail as a separate dump instead of relying on simulator reverse-debug options. Finally, in a live 'watch only visible rows' mode, signals scrolled out of view have no data for that interval: show them as X/unknown, not as the last value, and regenerate detail by re-simulating from the nearest checkpoint. Verify that the re-simulation is deterministic, especially on testbenches with races.