[thomdehoog]

A low weighted score previously could leave forbidden motifs in the output while the interface implied that all requirements had been met. Mandatory requirements now take priority over soft preferences, and an independent final validator returns explicit validity and unresolved regions. A bounded coordinated repair phase can escape single-codon traps without claiming that failed heuristic searches prove a design impossible.

The algorithm also now:

• Uses length-stable global GC penalties and cools annealing on every iteration.
• Preserves the original input baseline across restarts and reports the strongest hairpins across the full construct.
• Supports fixed flanks, protected codons, explicit codon exclusions, and selectable mandatory criteria.
• Makes splice screening opt-in, keeps automatic host selection conservative, and exposes strand policy.
• Records the effective seed, search settings, table content/hash, backend metadata, and full configuration for replay.

This is a 0.3.0 behavior change: motif avoidance is mandatory by default, splice screening defaults to none, and optimized sequences/scores may differ from earlier versions. The interface changes show validity accurately and account for construct coordinates.

Validation:

• Dependency-free Python suite: 246 passed, 2 skipped.
• With the MaxEntScan source backend: 247 passed, 1 skipped.
• 3,000 randomized synonymous edits checked incremental scores and mandatory counts against full recomputation.
• Frontend type-check, committed-bundle verification, and build passed.
• All GitHub CI jobs passed: Python 3.10–3.13, the MaxEntScan job, and all 13 Playwright browser tests, including unresolved-design reporting and cancellation.
• Local Chromium downloads were unavailable; browser coverage was verified on GitHub Actions.

Repair remains a bounded heuristic. Validity applies to the configured sequence criteria; these checks do not establish experimental expression or replace thermodynamic RNA modelling.