Eviction loop over an ordered set goes O(N^2) when every candidate is ineligible at the same timestamp

Pattern: a bounded-memory writer keeps per-signal pending buffers in a std::set ordered by size and, when a cap is hit, calls evictOne(t) in a loop: walk the set from the largest entry, skip entries whose last row is at the current time t (they cannot be flushed yet), flush the first eligible one. Normal runs are fine because most entries are older than t. But a snapshot of every signal at one timestamp (initial or final dump) makes every entry ineligible: each change() walks the whole set, finds nothing and falls through, so N changes cost N^2/2 set steps (262K signals: >10 min of CPU, looked like a hang, with flat memory).\n\nFix that keeps the same choices (byte-identical output): remember a scan that found nothing at time t (noEvict=true, noEvictT=t) and return false at once for further calls at the same t. Clear the memo wherever an entry can become eligible again at the same t. Here that was only the block-close path, which leaves an empty, still-reserved entry. Measured: 10,670 vs 846 set steps per value at 20K signals, flat at ~1K/value up to 80K.\n\nTesting tips:\n- Add a scan counter to the stats and assert steps-per-value below a bound, instead of timing.\n- Keep the old scan behind a -D..._RED macro so the same harness builds red and green and compares output hashes.\n- Comparing against a run with a larger cap is the wrong oracle: a binding cap legitimately changes block boundaries, so the bytes differ for reasons unrelated to the fix.\n\nSide gotchas from the same session:\n- git archive <ref> run from a subdirectory archives only that subtree, so the extracted tree has no leading src/ directory.\n- When a sed replacement contains a Verilog delay (#500), a # delimiter breaks it. Use perl -pe with the text in an env var: X="..." perl -pe "s/^(pat)/$ENV{X}\\n$1/".