Fix Ukrainian translation quality across OncoWiki (render engine, Onco Wiki, site, KB content)

[romeo111]

Summary

User audit ("клацнув по сайту — переклад поганий, незрозумілий") found that Ukrainian-language pages across openonco.info were frequently showing raw English text. Root cause traced to three independent bug classes, fixed across the whole stack, plus a full-corpus content-quality pass and two follow-up schema fixes.

Render engine (knowledge_base/engine/render.py, plan.py, _actionability.py, _translation_overrides.py, schemas/plan.py)

• The render layer produces Ukrainian by default (English is derived via post-hoc string substitution), so any hardcoded English in the default path leaked onto every treatment plan. ~60 confirmed instances fixed:
  • _UI_STRINGS had ~34 entries whose "uk" value was itself raw English (e.g. monitoring/timeline/sources_cited).
  • Several call sites (primary/alternative track headings, dt labels for Indication/Regimen/Reason/Supportive care, MDT talk-tree, evidence-lane labels, the no-plan-generated fallback page) hardcoded English literals that bypassed the _t() lookup entirely.
  • VariantActionabilityHit.evidence_summary had no schema channel for a translated companion to reach the render layer at all — added evidence_summary_ua.
  • track.selection_reason sentences were hardcoded English templates in plan.py; rewritten to author Ukrainian by default with EN mirrors registered in _translation_overrides.py.
  • Regimen names showed the raw English name instead of name_ua in four places.

Onco Wiki (scripts/build_kb_wiki.py)

• Same bug class in a separate code path (loads YAML directly, not through the engine): 463 biomarker-actionability pages and 636 red-flag pages were silently ignoring their _ua companion field and showing English regardless of locale.
• A literal Python str(True)/str(False) was leaking as the word "True" into the "Registered in Ukraine" column instead of "Так"/"Ні".

Site chrome + examples (scripts/build_site.py, scripts/site_cases.py)

• 15 spots where an English literal survived while sibling strings in the same list/dict were already translated (About page title/h1, try.html buttons/errors, specs.html labels, landing carousel).
• 286 of 295 case-gallery card titles had label_ua byte-identical to label_en — never translated (e.g. "NSCLC - ALK fusion - ESCAT IA / CIViC evidence" shown verbatim). Translated the ordinary connector words while preserving gene symbols/drug INNs/trial acronyms/disease-abbreviation shorthand in Latin script per the project's stay_latin glossary convention.

KB content-quality sweep (knowledge_base/hosted/content/**)

• Parallel review across every entity type (diseases, biomarkers, biomarker_actionability, redflags, regimens, drugs, indications, algorithms, procedures, radiation_courses): 2,982 files checked, 1,919 edited. Fixed EN bleed, garbled/mixed-script transliterations, drug-name misspellings, grammar agreement errors, dropped content, and glossary-locked-phrase substitutions — while preserving gene symbols, drug INNs, trial acronyms, and classification codes in Latin script per knowledge_base/translation/ua_terminology.yaml.
• All edited entities keep ukrainian_review_status: pending_clinical_signoff with draft_revision bumped — a language-quality pass, not a clinical content change.

Two follow-up schema fixes (surfaced while investigating pre-existing test failures)

1. BiomarkerClinicalContext enum — 15 prevention-persona screening biomarker files consistently used a richer 8-value taxonomy than the ratified 5-value enum. Extended the enum with the 4 missing values (content was ahead of the schema, not wrong) and dropped a redundant duplicate token. Resolved all 15 pre-existing KB validator schema errors, which had cascaded into 59 test failures.
2. MonitoringSchedule — the one entity type with zero schema support for a Ukrainian companion field despite the render layer displaying its content on every plan's Monitoring/Timeline sections. Added window_ua/checkpoints_ua/visits_ua/notes_ua, wired the renderer to prefer them, and translated all 20 monitoring-schedule files.

Test plan

• Full pytest suite: 2337 passed, 23 failed, 9 skipped, 0 errors (down from a confirmed pre-existing baseline of 40 failed / 85 errors before the schema fix). All 23 remaining failures independently confirmed pre-existing and unrelated to this change (bundle-size ceilings, a Windows bash path-separator issue in test_tasktorrent_scripts_smoke.py, stale date assertions in test_ukraine_registration.py, a redflag coverage baseline) — verified via git stash isolation and cross-checking untouched files.
• knowledge_base/validation/ua_quality.py error count: 684 → 372 over the KB content sweep (EN_BLEED 351 → 50).
• Site rebuilt (scripts/build_site.py) and spot-checked rendered docs/ukr/ pages after each wave of fixes.
• KB loader/schema validation passes cleanly for all touched entity types.

Notes for reviewers

• claude/amazing-brattain-b7353b (unmerged) has an independently-authored render.py change for the same MonitoringSchedule UA feature (commit f035f3bd5d) — worth reconciling on next integration to avoid divergence.
• Translation content is machine-drafted and marked pending_clinical_signoff throughout, per CHARTER §6.1 / §8.3 — needs Clinical Co-Lead review before any further promotion, same as prior translation waves.
• .github/PULL_REQUEST_TEMPLATE.md is scoped to a different workflow (benchmark cancer-report submissions) and doesn't apply to this engineering change, so it wasn't used verbatim.

🤖 Generated with Claude Code