v0.3.4

Changed

• Improve: expand_query prompt now instructs the LLM to avoid standalone audience/demographic terms (children, family, professional, etc.), reducing false matches from contextual noise. Adds explicit rule 11 for AUDIENCE QUALIFIERS directing expansion to focus on the topic, not the audience.

Added

• indexer property test coverage. Added testIndexerDefaultsToAuto and testFromArrayMapsIndexer to ScoltaConfigTest — the indexer property (default 'auto', accepts 'php'/'binary'/'auto') had zero test coverage.
• AI feature toggle enforcement in AiEndpointHandler. When aiExpandQuery=false or aiSummarize=false, the handler now returns ['ok'=>false, 'status'=>404, 'error'=>'Feature disabled'] immediately without calling the AI service. AiControllerTrait::createHandler() passes these flags from ScoltaConfig. Added five tests: disabled returns 404, disabled does not call AI service (expand and summarize), follow-up unaffected by either toggle.
• AI configuration tests (Phase 5). Added testAiLanguagesFlagsPropagateToJsOutput: sets ai_languages, ai_expand_query=false, ai_summarize=false, max_follow_ups=0 and confirms all appear correctly in toJsScoringConfig() output.
• Display behavior tests (Phase 2). Added excerpt_length to testToJsScoringConfigValuesMatchConfig to confirm EXCERPT_LENGTH in JS output reflects config.
• Scoring behavior tests (Phase 1). ScoltaConfigTest: completeness check for all 25 toJsScoringConfig() keys, value-mapping assertions, phrase-proximity field assertions, and a negative test confirming server-side keys (cacheTtl, aiApiKey, etc.) are absent from the JS output. AiEndpointHandlerTest: testCacheTtlZeroNeverReadsCache, testCacheTtlZeroNeverWritesCache, testMaxFollowUpsZeroBlocksImmediately (with TrackingCacheDriver). New tests/Service/AiServiceAdapterTest.php: custom prompt overrides returned raw without {SITE_NAME} substitution; default prompts resolve site name and description; empty overrides fall back to default.

Fixed

• Hygiene: Replaced uniqid('', true) with bin2hex(random_bytes(8)) in IndexMerger — avoids period-containing directory names that can confuse cleanup scripts.
• Hygiene: Removed @ error suppression from mkdir calls in IndexMerger, PagefindFormatWriter, and StreamingFormatWriter; replaced with explicit is_dir() fallback + RuntimeException.
• Hygiene: Removed @unserialize in PhpIndexer; replaced with try/catch \Throwable so corrupt cache entries surface in logs instead of silently recomputing.
• Hygiene: Added === false error checks to all bare file_put_contents calls in ContentExporter, PagefindFormatWriter, and StreamingFormatWriter.
• Hygiene: Added TOCTOU-safe comments to intentional @unlink calls in BuildState.
• Hygiene: Added source-parse tests preventing reintroduction of @mkdir, @unlink outside BuildState, uniqid(..., true), unchecked file_put_contents, and unserialize without allowed_classes.
• Summarize and follow_up prompts now include a GROUNDING CHECK section. The new section
  instructs the LLM to verify each fact against the provided excerpts before citing it, extract
  whatever IS relevant from partially-matching excerpts, note any gaps, and suggest specific search
  terms — replacing the old binary "no results" fallback that could cause the LLM to hallucinate
  or discard partially-relevant content entirely.
• Summarize prompt now requires per-excerpt scanning and a minimum bullet count. The FORMAT
  RULES bullet was rewritten to instruct the LLM to scan each excerpt individually and produce
  at least 3-5 detail bullets when content is present, rather than conditionally adding "a
  bulleted list" only when details happen to be obvious. Reduces sparse or single-bullet responses
  for queries with multiple relevant excerpts.