Benchmarking a custom waveform store against the FSDB reader: use ffrSetLoadNavDBOnlyOnce or the comparison is unfair

The default FsdbReader reloads its navigation DB on every ffrLoadSignals, at roughly 110 ms per call on a 1.7 GB file. Comparisons against that default look like 10-600x speed-ups.

With ffrSetLoadNavDBOnlyOnce(1) set before the first load, a fixed signal cohort added once, and ResetViewWindow -> LoadSignals -> walk -> UnloadSignals per window, the reader does a point query on a stored clock in ~0.8 ms and 20 stored signals in ~1 ms.

Measured against that, a per-signal zstd block store (16K changes per block, no decoded cache) was: - slower on point queries for stored signals (1.3 vs 1.06 ms for 20 rows), because every query decodes whole blocks and a wide-struct block is ~1 MB; - 1.1-2.5x faster for computed signals and wide windows; - far faster on first touch.

Tips: - Bit-compare both sides with the same collapse rule: value at t0, then changes in (t0, t1], same timestamp last-wins, equal consecutive values dropped. With a view window, the reader returns an entry exactly at t0. - Time first touch (open + scope tree + first load) separately from steady state. - The manual says loading signals outside the cohort later, under the pinned navigation, can return wrong results. Keep cohorts fixed per process.