AI ingestion: dynamic role/category creation + tags, hardcoded read-only prompt

- Unknown roles from the AI are now resolved-or-CREATED (Persian-normalized dedupe) instead of dropped/fallback; new role gets the AI's category, assigned to the applicant.
- AI output gains category + tags; AI-detected skills/requirements (ICU, MMT, پروانه‌دار…) now fold into the applicant's searchable Tags.
- System prompt is hardcoded in AppSetting.DefaultPrompt and used directly by the auditor; admin sees it read-only (cannot edit/break it).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/JobsMedical.Web/Services/Scraping/AiAuditor.cs

@@ -9,7 +9,10 @@ namespace JobsMedical.Web.Services.Scraping;
 public record AiStructured(
     string? Kind, string? Role, string? City, string? District, string? ShiftType,
     string? EmploymentType, long? PayAmount, int? SharePercent, string? Title, string? FacilityName,
-    string? Phone = null, string? PersonName = null, int? YearsExperience = null, bool? IsLicensed = null);
+    string? Phone = null, string? PersonName = null, int? YearsExperience = null, bool? IsLicensed = null,
+    // Dynamic taxonomy: the model may name a role/category outside the seeded set (ingestion
+    // resolves-or-creates it). Tags carry the post's skills/requirements (ICU, MMT, پروانه‌دار…).
+    string? Category = null, IReadOnlyList<string>? Tags = null);
 
 /// <summary>An AI verdict on a raw listing.</summary>
 public record AiAuditResult(string Decision, int Confidence, string? Reason, AiStructured? Data)
@@ -45,7 +48,9 @@ public class OpenAiCompatibleAuditor : IAiAuditor
         confidence: عدد ۰ تا ۱۰۰
         reason: توضیح کوتاه فارسی
         kind: shift (شیفت توسط مرکز) | job (استخدام توسط مرکز) | talent (کادر درمان که خودش «آماده به کار» است)
-        role: عنوان دقیق نقش درمانی (مثل پرستار، پزشک عمومی، دندانپزشک، تکنسین اتاق عمل، ماما، کارشناس آزمایشگاه)
+        role: عنوان دقیق نقش درمانی (مثل پرستار، پزشک عمومی، دندانپزشک، تکنسین اتاق عمل، ماما، کارشناس آزمایشگاه). اگر تخصص دقیق در فهرست نبود، همان عنوان دقیق را برگردان.
+        category: گروه نقش (پزشک | پرستار | ماما | تکنسین | دندانپزشک). اگر هیچ‌کدام مناسب نبود، یک گروه کوتاه و مناسب پیشنهاد بده.
+        tags: آرایه‌ای از کلیدواژه‌های مهارت/الزام مرتبط به‌صورت رشته (مثل "ICU"، "MMT"، "CPR"، "پروانه‌دار"، "خانم") یا []
         city, district: نام شهر و محله/منطقه در صورت ذکر
         shiftType: day|evening|night|oncall (فقط برای shift)
         employmentType: fulltime|parttime|contract|plan
@@ -150,9 +155,9 @@ public class OpenAiCompatibleAuditor : IAiAuditor
             response_format = new { type = "json_object" },
             messages = new object[]
             {
-                // Admin prompt + an authoritative output schema, so classification/tags stay
-                // correct even if the stored prompt predates the talent/phone fields.
-                new { role = "system", content = s.AiSystemPrompt + "\n\n" + OutputSchema },
+                // Hardcoded, code-owned prompt (NOT the stored AiSystemPrompt) + the authoritative
+                // output schema, so classification/tags can never be broken by an admin edit.
+                new { role = "system", content = AppSetting.DefaultPrompt + "\n\n" + OutputSchema },
                 new { role = "user", content = "آگهی خام:\n" + rawText + "\n\nفقط با JSON پاسخ بده." },
             },
         };
@@ -214,11 +219,23 @@ public class OpenAiCompatibleAuditor : IAiAuditor
         long? L(string k) => r.TryGetProperty(k, out var v) && v.ValueKind == JsonValueKind.Number && v.TryGetInt64(out var n) ? n : null;
         int? NI(string k) => r.TryGetProperty(k, out var v) && v.ValueKind == JsonValueKind.Number && v.TryGetInt32(out var n) ? n : null;
         bool? B(string k) => r.TryGetProperty(k, out var v) && (v.ValueKind == JsonValueKind.True || v.ValueKind == JsonValueKind.False) ? v.GetBoolean() : null;
+        // Array-of-strings reader (tolerates the model returning a single string instead of an array).
+        IReadOnlyList<string>? SA(string k)
+        {
+            if (!r.TryGetProperty(k, out var v)) return null;
+            var list = new List<string>();
+            if (v.ValueKind == JsonValueKind.Array)
+                foreach (var el in v.EnumerateArray())
+                    if (el.ValueKind == JsonValueKind.String && el.GetString() is { Length: > 0 } s) list.Add(s);
+            else if (v.ValueKind == JsonValueKind.String && v.GetString() is { Length: > 0 } one) list.Add(one);
+            return list.Count > 0 ? list : null;
+        }
 
         var decision = (S("decision") ?? "review").ToLowerInvariant();
         var data = new AiStructured(S("kind"), S("role"), S("city"), S("district"), S("shiftType"),
             S("employmentType"), L("payAmount"), NI("sharePercent"), S("title"), S("facilityName"),
-            Phone: S("phone"), PersonName: S("personName"), YearsExperience: NI("yearsExperience"), IsLicensed: B("isLicensed"));
+            Phone: S("phone"), PersonName: S("personName"), YearsExperience: NI("yearsExperience"), IsLicensed: B("isLicensed"),
+            Category: S("category"), Tags: SA("tags"));
         return new AiAuditResult(decision, Math.Clamp(I("confidence", 50), 0, 100), S("reason"), data);
         }
     }

src/JobsMedical.Web/Services/Scraping/IngestionService.cs

@@ -166,14 +166,21 @@ public class IngestionService
 
         // One ad can name several roles («پرستار سالمند و کودک و همراه بیمار») — resolve them all
         // and publish one listing per role so each is browsable/filterable. Capped to avoid spam.
-        var roleNames = new List<string>();
-        if (!string.IsNullOrWhiteSpace(d?.Role)) roleNames.Add(d!.Role!.Trim());
-        roleNames.AddRange(parsed.RoleNames);
-        if (parsed.RoleName is not null) roleNames.Add(parsed.RoleName);
-        var pubRoles = roleNames
-            .Select(n => roles.FirstOrDefault(r => r.Name == n))
-            .Where(r => r is not null).Cast<Role>()
-            .Distinct().Take(4).ToList();
+        // The AI's role (+ its category) is the trusted, possibly-new one; parser names are already
+        // canonical matches. Unknown roles are CREATED (dynamic taxonomy), not dropped.
+        var candidates = new List<(string name, string? category)>();
+        if (!string.IsNullOrWhiteSpace(d?.Role)) candidates.Add((d!.Role!.Trim(), d.Category));
+        foreach (var n in parsed.RoleNames) candidates.Add((n, null));
+        if (parsed.RoleName is not null) candidates.Add((parsed.RoleName, null));
+
+        var pubRoles = new List<Role>();
+        foreach (var (name, category) in candidates)
+        {
+            if (string.IsNullOrWhiteSpace(name)) continue;
+            var role = ResolveOrCreateRole(roles, name, category);
+            if (!pubRoles.Contains(role)) pubRoles.Add(role);
+            if (pubRoles.Count >= 4) break;
+        }
         if (pubRoles.Count == 0) pubRoles.Add(roles.First());
 
         var city = cities.FirstOrDefault(c => c.Name == cityName)
@@ -205,7 +212,7 @@ public class IngestionService
                     Description = raw.RawText,
                     Status = ShiftStatus.Open, Source = ShiftSource.Aggregated, SourceUrl = raw.SourceUrl,
                     Contacts = BuildContacts(d, parsed),   // fresh instances per listing
-                    Tags = BuildTags(parsed, role, city),
+                    Tags = BuildTags(parsed, d, role, city),
                 });
             raw.Status = RawListingStatus.Normalized;
             return;
@@ -261,13 +268,52 @@ public class IngestionService
         raw.Status = RawListingStatus.Normalized;
     }
 
-    /// <summary>Space-separated searchable tags: parsed cert/skill tags + this listing's role + city.</summary>
-    private static string BuildTags(ParsedListing parsed, Role role, City city)
+    /// <summary>Space-separated searchable tags: parsed cert/skill tags + AI-detected skills/requirements
+    /// + this listing's role/category + city. Drives deep search and tag chips on the applicant card.</summary>
+    private static string BuildTags(ParsedListing parsed, AiStructured? d, Role role, City city)
     {
-        var tags = new List<string>(parsed.Tags) { role.Name, city.Name };
+        var tags = new List<string>(parsed.Tags) { role.Name, role.Category, city.Name };
+        if (d?.Tags is not null)
+            tags.AddRange(d.Tags.Where(t => !string.IsNullOrWhiteSpace(t)).Select(t => t.Trim()));
         return string.Join(" ", tags.Where(t => !string.IsNullOrWhiteSpace(t)).Distinct());
     }
 
+    /// <summary>Find an existing role by Persian-normalized name; if none, create a new Role (dynamic
+    /// taxonomy) using the AI's suggested category — reusing an existing category when one normalizes
+    /// to the same text — and add it to the in-run list so later items reuse it instead of duplicating.</summary>
+    private Role ResolveOrCreateRole(List<Role> roles, string name, string? category)
+    {
+        var norm = NormalizeFa(name);
+        var match = roles.FirstOrDefault(r => NormalizeFa(r.Name) == norm);
+        if (match is not null) return match;
+
+        var wantCat = string.IsNullOrWhiteSpace(category) ? "سایر" : category!.Trim();
+        // Collapse onto an existing category that normalizes the same, so «تکنسین» != «تکنسين» doesn't fork.
+        var existingCat = roles.Select(r => r.Category)
+            .FirstOrDefault(c => !string.IsNullOrWhiteSpace(c) && NormalizeFa(c) == NormalizeFa(wantCat));
+
+        var created = new Role
+        {
+            Name = Clamp(name.Trim(), 100),            // respect Role.Name MaxLength(100)
+            Category = Clamp(existingCat ?? wantCat, 50),  // respect Role.Category MaxLength(50)
+            IsActive = true,
+            SortOrder = (roles.Count == 0 ? 0 : roles.Max(r => r.SortOrder)) + 1,
+        };
+        _db.Roles.Add(created);
+        roles.Add(created);   // reuse within this run (saved with the batch at end of source)
+        _log.LogInformation("Ingestion introduced new role «{Role}» (category «{Category}») from AI.",
+            created.Name, created.Category);
+        return created;
+    }
+
+    /// <summary>Normalize a Persian string for dedupe: unify Arabic/Persian ي→ی and ك→ک, drop ZWNJ,
+    /// collapse whitespace, trim, lowercase (so Latin tags like "ICU"/"icu" also match).</summary>
+    private static string NormalizeFa(string? s) => Regex.Replace(
+        (s ?? "").Replace('ي', 'ی').Replace('ك', 'ک').Replace('‌', ' ').Trim(),
+        @"\s+", " ").ToLowerInvariant();
+
+    private static string Clamp(string s, int max) => s.Length <= max ? s : s[..max].Trim();
+
     /// <summary>Fresh ContactMethod rows for one talent listing (parser contacts + AI phone).</summary>
     private static List<ContactMethod> BuildContacts(AiStructured? d, ParsedListing parsed)
     {

src/JobsMedical.Web/Services/Scraping/SettingsService.cs

@@ -31,8 +31,7 @@ public class SettingsService
         s.AiEndpoint = incoming.AiEndpoint?.Trim();
         s.AiApiKey = incoming.AiApiKey?.Trim();
         s.AiModel = incoming.AiModel?.Trim();
-        s.AiSystemPrompt = string.IsNullOrWhiteSpace(incoming.AiSystemPrompt)
-            ? AppSetting.DefaultPrompt : incoming.AiSystemPrompt;
+        s.AiSystemPrompt = AppSetting.DefaultPrompt;   // hardcoded & read-only — keep the column in sync
         s.AiAutoApprove = incoming.AiAutoApprove;
         s.AiUseProxy = incoming.AiUseProxy;
         // Channel scraping sources