[Alexander-Mitrofanov]

Per-seed diversity needs candidate selection before global ranking; filtering final tracebacks loses some seeds' best sites. This implements --outNumberPerSeed=K using the maintainer's decisions in #288.

Fixes #285. Supersedes the documentation checkpoint in #288.

• Keep each seed's best K distinct sites across merged regions/windows, deduplicate their union using the lowest-energy eligible conformation, then apply final energy, overlap and count restrictions without refilling quotas. --outPerRegion resets quotas per region combination.
• Cover S/H, S/M, X/H, X/M, X/R, X/K, B/H and seed-only S/X/P. MFE passes constrain the generating seed and retain coherent energy/traceback pairs. P/H and P/M rank and report each left-most anchor's site ensemble, before full-site aggregation.
• Preserve the existing execution path at the default zero; reject positive quotas with --noSeed, ignore the option for evaluation, and accept parameter files and values in [0,1000]. Zall, trackers and actual-pair probabilities retain their ordinary prediction semantics.
• Document behavior, library usage, compatibility and runtime in README, ChangeLog and the updated implementation note.

Validation: GCC 14 / ViennaRNA 2.7.2 release and debug make tests -j2 both passed 9/9 suites. The API suite passed 123 cases (529,850 release / 529,851 debug assertions), including exhaustive chain enumeration, noLP and cropped-coordinate oracles, anchor callback comparisons across the existing theory/random fixtures, heuristic/bulged ensembles and retained traceback energies. CLI regressions cover the mode matrix, explicit zero, invalid values, parameter files, evaluation, regions/windows, thread determinism, final restrictions, unchanged trackers and byte-identical probability files. Installation, a linked/running pkg-config consumer, five affected public headers compiled independently, and git diff --check also passed.

Out-of-tree validation used the source src/ include path, the generated source-side IntaRNA.pc expected by current build rules, and unchanged CLI fixtures copied into each build's tests/data/; no expected outputs were regenerated.

Tradeoff: enabling this option runs the selected MFE predictor once per admitted seed, plus one ordinary pass when statistics/trackers are needed. Runtime can increase substantially; retained candidate storage is bounded by K sites per seed. Model P uses its existing anchor decomposition directly, including its canonical ownership of nested explicit seeds and the existing legacy bulged-mode limitations. Custom MFE predictor subclasses can override the per-seed factory hook; incompatible inherited factories fail explicitly instead of silently dropping custom recurrence behavior.

[martin-raden]

@Alexander-Mitrofanov

thanks, please extend:

incorporate into "gui"

• add a new input field for "--outNumberPerSeed" similar to "outNumber" input on right of "outNumber" and left of "outOverlap"
  • default value = "0"
• move "outMaxE", "outDeltaE" and "outMinPu" into a new second row (making the current 2nd row the 3rd)