Fix filtering config save, auto-save, streamed results, improved AI prompt

- Add missing fields to FilteringConfigSchema (aiParseFiles, startupAdvanceCount,
  conceptAdvanceCount, notifyOnEntry, notifyOnAdvance) — Zod was silently
  stripping them on save
- Restore auto-save with 800ms debounce on config changes
- Add staggered animations for filtering results (stream in one-by-one)
- Improve AI screening prompt: file type label mappings, soft cap handling,
  missing documents = fail, better user prompt structure

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/server/services/ai-filtering.ts

@@ -149,7 +149,7 @@ const MAX_PARALLEL_BATCHES = 10
 const AI_SCREENING_SYSTEM_PROMPT = `You are an expert project screening assistant for an ocean conservation competition.
 
 ## Your Role
-Evaluate each project against the provided screening criteria. Be objective and base evaluation only on provided data.
+Evaluate each project against the provided screening criteria. Be objective and base evaluation only on provided data. Your goal is to identify projects that clearly FAIL the criteria — not to nitpick minor issues.
 
 ## Output Format
 Return a JSON object with this exact structure:
@@ -179,12 +179,43 @@ Return a JSON object with this exact structure:
 - founded_year: when the company/initiative was founded (use for age checks)
 - ocean_issue: the ocean conservation area
 - file_count, file_types: uploaded documents summary
-- files[]: per-file details with file_type, page_count (if known), size_kb, detected_lang (ISO 639-3 language code like 'eng', 'fra'), lang_confidence (0-1), round_name (which round the file was submitted for), and is_current_round flag
+- files[]: per-file details (see File Type Reference below) with file_type, page_count (if known), size_kb, detected_lang (ISO 639-3 language code like 'eng', 'fra'), lang_confidence (0-1), round_name (which round the file was submitted for), and is_current_round flag
 - description: project summary text
 - tags: topic tags
 - If document content is provided (text_content field in files), use it for deeper analysis. Pay SPECIAL ATTENTION to files from the current round (is_current_round=true) as they are the most recent and relevant submissions.
 - If detected_lang is provided, use it to evaluate language requirements (e.g. 'eng' = English, 'fra' = French). lang_confidence indicates detection reliability.
 
+## File Type Reference
+The file_type field uses these codes. When criteria mention document names, match them to the correct code:
+- EXEC_SUMMARY = Executive Summary
+- PRESENTATION = Pitch Deck / Presentation
+- BUSINESS_PLAN = Business Plan
+- VIDEO = Video
+- VIDEO_PITCH = Video Pitch
+- SUPPORTING_DOC = Supporting Document
+- OTHER = Other / miscellaneous document
+
+IMPORTANT: Only evaluate page limits against the CORRECT document type. For example, a page limit on "executive summary" applies ONLY to files with file_type=EXEC_SUMMARY, not to pitch decks or other documents.
+
+## How to Interpret Criteria
+- Treat criteria as reasonable guidelines written by a human, not as rigid legal rules.
+- When criteria say "soft cap", "approximately", "around", "about", or "reasonable amount", be LENIENT. A document that is 1-3 pages over a soft cap still meets criteria. Only flag if it is egregiously over (e.g. double the stated limit).
+- When criteria have NO softening language, treat them as firm but still use judgment — a 1-page overage on a hard limit is borderline, not an automatic fail.
+- Page count data may be unavailable (null) for some files. If page_count is null, do NOT penalize — only evaluate page limits when page_count is actually provided for that file.
+- Focus on the INTENT behind each criterion, not a hyper-literal reading.
+
+## Missing Documents
+- If criteria mention document requirements (executive summary, pitch deck, etc.) and the project has ZERO files (file_count=0), this is a CLEAR FAIL. Every competition applicant is expected to upload documents.
+- "Where applicable" in criteria refers to edge cases within specific document types — it does NOT mean documents are optional.
+- A project with no uploaded documents at all should always have meets_criteria=FALSE unless the criteria explicitly say documents are optional.
+
+## Decision Framework
+- Set meets_criteria=TRUE if the project satisfies the core intent of the criteria, even with minor imperfections (e.g. a few pages over a soft cap).
+- Set meets_criteria=FALSE when there is a clear, material violation: missing required documents entirely, startup far exceeding age limit, documents entirely in the wrong language, no visible ocean impact, etc.
+- When in doubt on a SOFT criterion (page lengths, formatting), set meets_criteria=TRUE with a lower confidence score and note the concern in reasoning.
+- When in doubt on a HARD criterion (age limits, language, having documents at all), set meets_criteria=FALSE.
+- Mention specific concerns in the reasoning field so the admin can review.
+
 ## Guidelines
 - Evaluate ONLY against the provided criteria, not your own standards
 - A confidence of 1.0 means absolute certainty; 0.5 means borderline
@@ -192,7 +223,7 @@ Return a JSON object with this exact structure:
 - When criteria differ by category (e.g. stricter for STARTUP vs BUSINESS_CONCEPT), apply the appropriate threshold
 - When criteria mention regional considerations (e.g. African projects), use the country/region fields
 - Do not include any personal identifiers in reasoning
-- If project data is insufficient to evaluate, set confidence below 0.3`
+- If project data is insufficient to evaluate, set confidence below 0.3 and default meets_criteria to TRUE`
 
 // ─── Field-Based Rule Evaluation ────────────────────────────────────────────
 
@@ -371,10 +402,16 @@ async function processAIBatch(
   // Sanitize user-supplied criteria
   const { sanitized: safeCriteria } = sanitizeUserInput(criteriaText)
 
-  // Build optimized prompt
-  const userPrompt = `CRITERIA: ${safeCriteria}
-PROJECTS: ${JSON.stringify(anonymized)}
-Evaluate and return JSON.`
+  // Build user prompt with clear structure
+  const userPrompt = `## Screening Criteria
+The admin has defined the following requirements. Evaluate each project against ALL of these criteria:
+
+${safeCriteria}
+
+## Projects to Evaluate (${anonymized.length} total)
+${JSON.stringify(anonymized)}
+
+Evaluate each project and return JSON with your assessment.`
 
   const MAX_PARSE_RETRIES = 2
   let parseAttempts = 0