Divar geo-coords to facility map + medical gate + RawListing FK/geo migrations

src/JobsMedical.Web/Data/AppDbContext.cs

@@ -171,5 +171,14 @@ public class AppDbContext : DbContext, IDataProtectionKeyContext
         // Dedupe ingested listings by content hash.
         b.Entity<RawListing>().HasIndex(r => r.ContentHash);
         b.Entity<RawListing>().HasIndex(r => r.Status);
+        // A RawListing only LINKS to the post it produced — it must outlive that post (it's the
+        // dedupe cache). So deleting a Shift/Talent NULLs the back-reference rather than orphaning a
+        // dangling FK or blocking the delete. LinkedTalentId previously had no FK at all (orphan risk).
+        b.Entity<RawListing>()
+            .HasOne(r => r.LinkedShift).WithMany()
+            .HasForeignKey(r => r.LinkedShiftId).OnDelete(DeleteBehavior.SetNull);
+        b.Entity<RawListing>()
+            .HasOne(r => r.LinkedTalent).WithMany()
+            .HasForeignKey(r => r.LinkedTalentId).OnDelete(DeleteBehavior.SetNull);
     }
 }

src/JobsMedical.Web/Migrations/20260609173744_RawListingLinkFks.cs

@@ -0,0 +1,69 @@
+using Microsoft.EntityFrameworkCore.Migrations;
+
+#nullable disable
+
+namespace JobsMedical.Web.Migrations
+{
+    /// <inheritdoc />
+    public partial class RawListingLinkFks : Migration
+    {
+        /// <inheritdoc />
+        protected override void Up(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.DropForeignKey(
+                name: "FK_RawListings_Shifts_LinkedShiftId",
+                table: "RawListings");
+
+            // LinkedTalentId never had an FK before, so existing rows may point at deleted talent.
+            // Null those orphans first, otherwise AddForeignKey below fails on a populated DB.
+            migrationBuilder.Sql(
+                "UPDATE \"RawListings\" r SET \"LinkedTalentId\" = NULL " +
+                "WHERE r.\"LinkedTalentId\" IS NOT NULL " +
+                "AND NOT EXISTS (SELECT 1 FROM \"TalentListings\" t WHERE t.\"Id\" = r.\"LinkedTalentId\");");
+
+            migrationBuilder.CreateIndex(
+                name: "IX_RawListings_LinkedTalentId",
+                table: "RawListings",
+                column: "LinkedTalentId");
+
+            migrationBuilder.AddForeignKey(
+                name: "FK_RawListings_Shifts_LinkedShiftId",
+                table: "RawListings",
+                column: "LinkedShiftId",
+                principalTable: "Shifts",
+                principalColumn: "Id",
+                onDelete: ReferentialAction.SetNull);
+
+            migrationBuilder.AddForeignKey(
+                name: "FK_RawListings_TalentListings_LinkedTalentId",
+                table: "RawListings",
+                column: "LinkedTalentId",
+                principalTable: "TalentListings",
+                principalColumn: "Id",
+                onDelete: ReferentialAction.SetNull);
+        }
+
+        /// <inheritdoc />
+        protected override void Down(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.DropForeignKey(
+                name: "FK_RawListings_Shifts_LinkedShiftId",
+                table: "RawListings");
+
+            migrationBuilder.DropForeignKey(
+                name: "FK_RawListings_TalentListings_LinkedTalentId",
+                table: "RawListings");
+
+            migrationBuilder.DropIndex(
+                name: "IX_RawListings_LinkedTalentId",
+                table: "RawListings");
+
+            migrationBuilder.AddForeignKey(
+                name: "FK_RawListings_Shifts_LinkedShiftId",
+                table: "RawListings",
+                column: "LinkedShiftId",
+                principalTable: "Shifts",
+                principalColumn: "Id");
+        }
+    }
+}

src/JobsMedical.Web/Migrations/20260609174915_RawListingGeo.cs

@@ -0,0 +1,38 @@
+using Microsoft.EntityFrameworkCore.Migrations;
+
+#nullable disable
+
+namespace JobsMedical.Web.Migrations
+{
+    /// <inheritdoc />
+    public partial class RawListingGeo : Migration
+    {
+        /// <inheritdoc />
+        protected override void Up(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.AddColumn<double>(
+                name: "Lat",
+                table: "RawListings",
+                type: "double precision",
+                nullable: true);
+
+            migrationBuilder.AddColumn<double>(
+                name: "Lng",
+                table: "RawListings",
+                type: "double precision",
+                nullable: true);
+        }
+
+        /// <inheritdoc />
+        protected override void Down(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.DropColumn(
+                name: "Lat",
+                table: "RawListings");
+
+            migrationBuilder.DropColumn(
+                name: "Lng",
+                table: "RawListings");
+        }
+    }
+}

src/JobsMedical.Web/Migrations/AppDbContextModelSnapshot.cs

@@ -748,12 +748,18 @@ namespace JobsMedical.Web.Migrations
                     b.Property<DateTime>("FetchedAt")
                         .HasColumnType("timestamp with time zone");
 
+                    b.Property<double?>("Lat")
+                        .HasColumnType("double precision");
+
                     b.Property<int?>("LinkedShiftId")
                         .HasColumnType("integer");
 
                     b.Property<int?>("LinkedTalentId")
                         .HasColumnType("integer");
 
+                    b.Property<double?>("Lng")
+                        .HasColumnType("double precision");
+
                     b.Property<string>("ParsedJson")
                         .HasColumnType("text");
 
@@ -783,6 +789,8 @@ namespace JobsMedical.Web.Migrations
 
                     b.HasIndex("LinkedShiftId");
 
+                    b.HasIndex("LinkedTalentId");
+
                     b.HasIndex("Status");
 
                     b.ToTable("RawListings");
@@ -1415,9 +1423,17 @@ namespace JobsMedical.Web.Migrations
                 {
                     b.HasOne("JobsMedical.Web.Models.Shift", "LinkedShift")
                         .WithMany()
-                        .HasForeignKey("LinkedShiftId");
+                        .HasForeignKey("LinkedShiftId")
+                        .OnDelete(DeleteBehavior.SetNull);
+
+                    b.HasOne("JobsMedical.Web.Models.TalentListing", "LinkedTalent")
+                        .WithMany()
+                        .HasForeignKey("LinkedTalentId")
+                        .OnDelete(DeleteBehavior.SetNull);
 
                     b.Navigation("LinkedShift");
+
+                    b.Navigation("LinkedTalent");
                 });
 
             modelBuilder.Entity("JobsMedical.Web.Models.Review", b =>

src/JobsMedical.Web/Models/RawListing.cs

@@ -25,10 +25,16 @@ public class RawListing
     public Shift? LinkedShift { get; set; }
 
     public int? LinkedTalentId { get; set; }          // آگهی «آماده به کار» ساخته‌شده از این متن
+    public TalentListing? LinkedTalent { get; set; }
 
     [MaxLength(500)]
     public string? SourceUrl { get; set; }
 
+    /// <summary>Approximate coordinates harvested from the source (e.g. Divar's fuzzed map center).
+    /// Carried through the review queue so a manual publish can still place the facility on the map.</summary>
+    public double? Lat { get; set; }
+    public double? Lng { get; set; }
+
     /// <summary>SHA-256 of the normalized text — used to dedupe across ingestion runs.</summary>
     [MaxLength(64)]
     public string? ContentHash { get; set; }

src/JobsMedical.Web/Pages/Admin/Index.cshtml

@@ -40,6 +40,14 @@
             <p class="muted" style="font-size:11px; margin:8px 0 0;">
                 موتور: واکشی ← حذف تکراری ← تجزیه ← اعتبارسنجی ← صف بررسی.
             </p>
+            <form method="post" onsubmit="return confirm('⚠ همه‌ی آیتم‌های جمع‌آوری‌شده (کش) و همه‌ی آگهی‌های منتشرشده از جمع‌آوری حذف می‌شوند (آگهی‌های ثبت‌شده توسط مراکز دست‌نخورده می‌مانند)، سپس همه‌چیز با هوش مصنوعی دوباره جمع‌آوری و افزوده می‌شود. این کار بازگشت‌ناپذیر است. ادامه می‌دهی؟');">
+                <button type="submit" asp-page-handler="PurgeAndReingest" class="btn btn-outline btn-block" style="margin-top:8px; color:var(--danger); border-color:var(--danger);">
+                    🔄 پاک‌سازی کش و جمع‌آوری مجدد با هوش مصنوعی
+                </button>
+            </form>
+            <p class="muted" style="font-size:11px; margin:6px 0 0;">
+                کش حذف تکراری و آگهی‌های جمع‌آوری‌شده پاک و از نو با AI پردازش می‌شوند. (آگهی‌های مراکز حذف نمی‌شوند.)
+            </p>
 
             <hr style="border:none; border-top:1px solid var(--line); margin:16px 0;" />
 

src/JobsMedical.Web/Pages/Admin/Index.cshtml.cs

@@ -65,6 +65,35 @@ public class IndexModel : PageModel
         return RedirectToPage();
     }
 
+    /// <summary>
+    /// DESTRUCTIVE rebuild, in two distinct deletes:
+    ///   1. The DEDUPE CACHE — ALL RawListings, including any added via «افزودن دستی». These are not
+    ///      published content; they're the crawl/staging rows whose ContentHash blocks re-ingesting
+    ///      the same ad. Wiping them lets everything be re-fetched and re-judged by the AI.
+    ///   2. AGGREGATED listings only — Shifts/JobOpenings/TalentListings with Source==Aggregated, i.e.
+    ///      produced by ingestion. Employer/admin-posted listings (Source==Direct) are left untouched.
+    /// Then re-fetch everything and re-run it through the (now AI-enabled) pipeline.
+    /// RawListings are deleted first so their LinkedShift/LinkedTalent FKs (SetNull) don't dangle;
+    /// DB cascade clears ContactMethods / Applications / InterestEvents when the posts are deleted.
+    /// </summary>
+    public async Task<IActionResult> OnPostPurgeAndReingestAsync()
+    {
+        int rawCount, shifts, jobs, talent;
+        await using (var tx = await _db.Database.BeginTransactionAsync())
+        {
+            rawCount = await _db.RawListings.ExecuteDeleteAsync();                                   // clear dedupe cache
+            shifts = await _db.Shifts.Where(s => s.Source == ShiftSource.Aggregated).ExecuteDeleteAsync();
+            jobs = await _db.JobOpenings.Where(j => j.Source == ShiftSource.Aggregated).ExecuteDeleteAsync();
+            talent = await _db.TalentListings.Where(t => t.Source == ShiftSource.Aggregated).ExecuteDeleteAsync();
+            await tx.CommitAsync();
+        }
+
+        var s = await _ingest.RunAsync();   // fresh fetch → AI audit → publish/queue
+        IngestMessage = $"پاک‌سازی شد (حذف: {rawCount} آیتم کش، {shifts} شیفت، {jobs} استخدام، {talent} آماده‌به‌کارِ جمع‌آوری‌شده). " +
+                        $"جمع‌آوری مجدد: {s.TotalPublished} منتشر، {s.TotalQueued} در صف، {s.TotalFlagged} پرچم، {s.TotalSpam} اسپم، {s.TotalDuplicates} تکراری.";
+        return RedirectToPage();
+    }
+
     private async Task LoadAsync()
     {
         Queue = await _db.RawListings

src/JobsMedical.Web/Pages/Admin/Review.cshtml.cs

@@ -282,13 +282,26 @@ public class ReviewModel : PageModel
         if (cityId is null) return null;   // no cities seeded — cannot create a facility
 
         // No facility named in the ad → use/create the shared placeholder.
-        var name = string.IsNullOrWhiteSpace(NewFacilityName) ? UnknownFacilityName : NewFacilityName.Trim();
+        var isPlaceholder = string.IsNullOrWhiteSpace(NewFacilityName);
+        var name = isPlaceholder ? UnknownFacilityName : NewFacilityName.Trim();
+
+        // Approximate coords carried from the crawl (e.g. Divar). NEVER apply them to the shared
+        // «نامشخص» placeholder — it's reused across many ads, so a single ad's point would mislead.
+        bool HasGeo() => !isPlaceholder && Raw?.Lat is not null;
 
         // Reuse an existing facility that's exactly or closely the same (Persian-aware fuzzy
         // match), so we don't create duplicates like «بیمارستان میلاد» vs «میلاد».
         var all = await _db.Facilities.ToListAsync();
         var match = FacilityMatcher.FindBest(all, name, cityId);
-        if (match is not null) return match.Id;
+        if (match is not null)
+        {
+            if (HasGeo() && match.Lat is null && match.Lng is null)   // backfill only, never overwrite
+            {
+                match.Lat = Raw!.Lat; match.Lng = Raw.Lng;
+                await _db.SaveChangesAsync();
+            }
+            return match.Id;
+        }
 
         var facility = new Facility
         {
@@ -297,6 +310,8 @@ public class ReviewModel : PageModel
             Type = FacilityType.Hospital,
             Verification = VerificationStatus.Unverified,
             IsVerified = false,
+            Lat = HasGeo() ? Raw!.Lat : null,
+            Lng = HasGeo() ? Raw!.Lng : null,
         };
         _db.Facilities.Add(facility);
         await _db.SaveChangesAsync();

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

@@ -59,17 +59,25 @@ public class DivarListingSource : IListingSource
                     continue;
                 }
                 using var doc = JsonDocument.Parse(body);
+                var cityLabel = CityLabel(s.DivarCity);   // every result is from the city we searched
                 foreach (var (text, token) in Harvest(doc.RootElement).Take(25))
                 {
                     var url = token is not null ? $"https://divar.ir/v/{token}" : "https://divar.ir";
-                    var withPhone = text;
+                    var itemText = text;
+                    // Stamp the city so the parser/AI always resolve a location (Divar's own location
+                    // line isn't always in the search row; the searched city is authoritative).
+                    if (!string.IsNullOrWhiteSpace(cityLabel) && !text.Contains(cityLabel))
+                        itemText += $"\n📍 {cityLabel}";
+                    double? lat = null, lng = null;
                     if (token is not null)
                     {
-                        var phones = await RevealPhonesAsync(client, token, s, ct);
-                        if (phones.Count > 0 && !phones.Any(text.Contains))
-                            withPhone = text + "\nشماره تماس: " + string.Join("، ", phones);
+                        // One detail fetch yields BOTH the phone and the map coordinates.
+                        var (phones, gLat, gLng) = await FetchDetailAsync(client, token, ct);
+                        if (phones.Count > 0 && !phones.Any(itemText.Contains))
+                            itemText += "\nشماره تماس: " + string.Join("، ", phones);
+                        lat = gLat; lng = gLng;
                     }
-                    items.Add(new ScrapedItem("دیوار", withPhone, url));
+                    items.Add(new ScrapedItem("دیوار", itemText, url, lat, lng));
                 }
             }
             catch (Exception ex) { _log.LogWarning(ex, "Divar fetch failed for query {Query}", q); }
@@ -95,16 +103,31 @@ public class DivarListingSource : IListingSource
         };
     }
 
+    /// <summary>Persian display name for the searched city (slug/number/Persian → Persian), used to
+    /// stamp every Divar result with its (authoritative) location.</summary>
+    private static string CityLabel(string? city) => (city ?? "").Trim().ToLowerInvariant() switch
+    {
+        "1" or "tehran" or "تهران" => "تهران",
+        "3" or "isfahan" or "esfahan" or "اصفهان" => "اصفهان",
+        "4" or "mashhad" or "مشهد" => "مشهد",
+        "5" or "shiraz" or "شیراز" => "شیراز",
+        "6" or "tabriz" or "تبریز" => "تبریز",
+        "1745" or "karaj" or "کرج" => "کرج",
+        _ => (city ?? "").Trim(),
+    };
+
     // The post detail endpoint returns the FULL description — many Divar job ads write the phone
     // straight into the body, so we can harvest it without Divar's (login-gated) contact reveal.
     private const string PostDetailUrl = "https://api.divar.ir/v8/posts-v2/web/";
 
     /// <summary>
-    /// Fetch a post's detail JSON and harvest any contact number it contains (mostly numbers the
-    /// poster wrote into the description). Divar's true "نمایش شماره" reveal is auth-gated; this
-    /// covers the common case where the number is in the ad text. Fails soft.
+    /// Fetch a post's detail JSON ONCE and harvest both (a) any contact number it contains (mostly
+    /// numbers the poster wrote into the description; Divar's true "نمایش شماره" reveal is auth-gated)
+    /// and (b) the post's APPROXIMATE map coordinates (the privacy-fuzzed center Divar shows as a
+    /// circle). Fails soft — returns whatever it could extract.
     /// </summary>
-    private async Task<List<string>> RevealPhonesAsync(HttpClient client, string token, AppSetting s, CancellationToken ct)
+    private async Task<(List<string> phones, double? lat, double? lng)> FetchDetailAsync(
+        HttpClient client, string token, CancellationToken ct)
     {
         try
         {
@@ -112,18 +135,68 @@ public class DivarListingSource : IListingSource
             req.Headers.TryAddWithoutValidation("User-Agent", Ua);
             req.Headers.TryAddWithoutValidation("Accept", "application/json");
             using var resp = await client.SendAsync(req, ct);
-            if (!resp.IsSuccessStatusCode) return new();
+            if (!resp.IsSuccessStatusCode) return (new(), null, null);
             var body = await resp.Content.ReadAsStringAsync(ct);
-            if (body.Contains("BLOCKING_VIEW")) return new();
-            return HtmlUtil.HarvestPhones(body);
+            if (body.Contains("BLOCKING_VIEW")) return (new(), null, null);
+            var phones = HtmlUtil.HarvestPhones(body);
+            double? lat = null, lng = null;
+            try { using var doc = JsonDocument.Parse(body); if (FindLatLng(doc.RootElement) is { } g) { lat = g.lat; lng = g.lng; } }
+            catch (JsonException) { /* detail wasn't JSON — phones still harvested from text */ }
+            return (phones, lat, lng);
         }
         catch (Exception ex)
         {
             _log.LogWarning(ex, "Divar detail/reveal failed for {Token}", token);
-            return new();
+            return (new(), null, null);
         }
     }
 
+    // Iran's bounding box — guards against picking up an unrelated number pair (timestamps, ids…).
+    private const double MinLat = 24, MaxLat = 40, MinLng = 44, MaxLng = 64;
+
+    /// <summary>
+    /// Tolerantly find an approximate (lat, lng) anywhere in Divar's detail JSON. Divar's shape
+    /// shifts (sometimes `latitude`/`longitude`, sometimes nested under `location`/`coordinates`),
+    /// so we walk the tree and accept the first OBJECT that holds BOTH a latitude-like and a
+    /// longitude-like numeric property whose values fall inside Iran. Pairing within one object
+    /// avoids matching a stray lat to an unrelated lng. Returns null if nothing plausible is found.
+    /// </summary>
+    private static (double lat, double lng)? FindLatLng(JsonElement el)
+    {
+        if (el.ValueKind == JsonValueKind.Object)
+        {
+            double? lat = null, lng = null;
+            foreach (var p in el.EnumerateObject())
+            {
+                if (lat is null && IsLatKey(p.Name) && TryNum(p.Value, out var la)) lat = la;
+                else if (lng is null && IsLngKey(p.Name) && TryNum(p.Value, out var lo)) lng = lo;
+            }
+            if (lat is double L && lng is double G && L is >= MinLat and <= MaxLat && G is >= MinLng and <= MaxLng)
+                return (L, G);
+            foreach (var p in el.EnumerateObject())
+                if (FindLatLng(p.Value) is { } r) return r;
+        }
+        else if (el.ValueKind == JsonValueKind.Array)
+            foreach (var item in el.EnumerateArray())
+                if (FindLatLng(item) is { } r) return r;
+        return null;
+    }
+
+    private static bool IsLatKey(string k) => k.Equals("latitude", StringComparison.OrdinalIgnoreCase) || k.Equals("lat", StringComparison.OrdinalIgnoreCase);
+    private static bool IsLngKey(string k) =>
+        k.Equals("longitude", StringComparison.OrdinalIgnoreCase) || k.Equals("lng", StringComparison.OrdinalIgnoreCase)
+        || k.Equals("lon", StringComparison.OrdinalIgnoreCase) || k.Equals("long", StringComparison.OrdinalIgnoreCase);
+
+    /// <summary>Coordinate may be a JSON number or a numeric string ("35.7"). Invariant culture.</summary>
+    private static bool TryNum(JsonElement v, out double d)
+    {
+        if (v.ValueKind == JsonValueKind.Number) return v.TryGetDouble(out d);
+        if (v.ValueKind == JsonValueKind.String)
+            return double.TryParse(v.GetString(), System.Globalization.NumberStyles.Float,
+                System.Globalization.CultureInfo.InvariantCulture, out d);
+        d = 0; return false;
+    }
+
     private static readonly string[] DescKeys =
         { "description", "middle_description_text", "subtitle", "bottom_description_text", "normal_text" };
 
@@ -134,9 +207,11 @@ public class DivarListingSource : IListingSource
             if (el.TryGetProperty("title", out var t) && t.ValueKind == JsonValueKind.String)
             {
                 var sb = new StringBuilder(t.GetString());
+                // Append ALL present description fields — the location/time line («… در تهران، جنت‌آباد»)
+                // is usually in bottom_description_text, so don't stop at the first match.
                 foreach (var k in DescKeys)
-                    if (el.TryGetProperty(k, out var d) && d.ValueKind == JsonValueKind.String)
-                    { sb.Append(" — ").Append(d.GetString()); break; }
+                    if (el.TryGetProperty(k, out var d) && d.ValueKind == JsonValueKind.String && d.GetString() is { Length: > 0 } v)
+                        sb.Append(" — ").Append(v);
                 var text = sb.ToString().Trim();
                 if (text.Length >= 15) yield return (text, FindToken(el));
             }

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

@@ -2,8 +2,11 @@ using JobsMedical.Web.Models;
 
 namespace JobsMedical.Web.Services.Scraping;
 
-/// <summary>One raw post pulled from a source (a Telegram message, a Divar ad, etc.).</summary>
-public record ScrapedItem(string Source, string RawText, string? SourceUrl = null);
+/// <summary>One raw post pulled from a source (a Telegram message, a Divar ad, etc.).
+/// Lat/Lng are an APPROXIMATE location when the source exposes one (e.g. Divar's privacy-fuzzed
+/// map center) — used to place an aggregated facility on the map / enable «near me».</summary>
+public record ScrapedItem(string Source, string RawText, string? SourceUrl = null,
+                          double? Lat = null, double? Lng = null);
 
 /// <summary>
 /// A pluggable source the ingestion engine pulls from. Configuration (enabled, channels, tokens)

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

@@ -46,6 +46,10 @@ public class IngestionService
 
     public IReadOnlyList<string> SourceNames => _sources.Select(s => s.Name).ToList();
 
+    /// <summary>Shared placeholder facility name for unnamed ads — kept identical to
+    /// Review.ResolveFacilityIdAsync so the auto-publish and manual-review flows reuse ONE record.</summary>
+    private const string UnknownFacilityName = "نامشخص / ثبت نشده";
+
     public async Task<IngestionSummary> RunAsync(CancellationToken ct = default)
     {
         var settings = await _settings.GetAsync();
@@ -71,7 +75,17 @@ public class IngestionService
             {
                 fetched++;
                 var hash = Hash(item.RawText);
-                if (await _db.RawListings.AnyAsync(r => r.ContentHash == hash, ct)) { dupes++; continue; }
+                var existing = await _db.RawListings.FirstOrDefaultAsync(r => r.ContentHash == hash, ct);
+                if (existing is not null)
+                {
+                    // Best-effort geo retry: coords are normally captured only on first ingest, but a
+                    // re-fetch may now expose a map center the first fetch lacked (Divar can fail-soft to
+                    // null on a bad response / out-of-bbox). Backfill the cached row when this fetch has
+                    // coords and the row has none, so an item still sitting in the queue can be placed on
+                    // the map when an admin publishes it. (A full refresh is the purge-and-reingest flow.)
+                    if (existing.Lat is null && item.Lat is not null) { existing.Lat = item.Lat; existing.Lng = item.Lng; }
+                    dupes++; continue;
+                }
 
                 var parsed = _parser.Parse(item.RawText, roleNames, cityNames, districtNames);
                 var val = _validator.Validate(item.RawText, parsed);
@@ -91,6 +105,7 @@ public class IngestionService
                     Confidence = confidence,
                     ValidationNotes = reason,
                     Status = status,
+                    Lat = item.Lat, Lng = item.Lng,   // approx. map coords (Divar) → facility on publish
                 };
                 _db.RawListings.Add(raw);
 
@@ -146,8 +161,15 @@ public class IngestionService
             var aiNote = Join($"AI: {ai.Decision} ({ai.Confidence}٪)" + (ai.Reason is null ? "" : $" — {ai.Reason}"), notes);
             if (ai.Reject) return (RawListingStatus.Discarded, aiNote, ai.Confidence);
             if (ai.Approve)
+            {
+                // MEDICAL GATE: the rule-validator's medical signal vetoes an AI approval. The AI can
+                // hallucinate (e.g. approved a GeekVape product ad 95% as a «پرستار» job) — when our
+                // own keyword/role check sees nothing clinical, never auto-publish; send to review.
+                if (!val.LooksMedical)
+                    return (RawListingStatus.Flagged, Join("هوش مصنوعی تأیید کرد ولی نشانهٔ کادر درمان یافت نشد — بررسی دستی", aiNote), ai.Confidence);
                 return (s.Mode == IngestionMode.Automatic && s.AiAutoApprove
                     ? RawListingStatus.Normalized : RawListingStatus.New, aiNote, ai.Confidence);
+            }
             return (RawListingStatus.Flagged, aiNote, ai.Confidence);  // review
         }
 
@@ -218,10 +240,15 @@ public class IngestionService
             return;
         }
 
-        // Never surface the crawl source (e.g. «مدجابز») in a public facility name.
+        // Never surface the crawl source (e.g. «مدجابز») in a public facility name. An unnamed ad
+        // falls back to ONE shared placeholder (same string as the manual-review flow, so both
+        // pipelines reuse a single record). That placeholder is shared by every unnamed ad in a
+        // city, so it must NEVER receive a single ad's fuzzy coords — that would mis-place dozens of
+        // unrelated listings on the map and in «near me». Mirrors Review.ResolveFacilityIdAsync.
+        bool unnamed = string.IsNullOrWhiteSpace(d?.FacilityName) && string.IsNullOrWhiteSpace(parsed.FacilityName);
         var facilityName = !string.IsNullOrWhiteSpace(d?.FacilityName) ? d!.FacilityName!.Trim()
             : !string.IsNullOrWhiteSpace(parsed.FacilityName) ? parsed.FacilityName!.Trim()
-            : "مرکز درمانی (نامشخص)";
+            : UnknownFacilityName;
         // Reuse an existing facility (exact or Persian-aware fuzzy match) before creating a new one.
         var facility = FacilityMatcher.FindBest(facilities, facilityName, city.Id);
         if (facility is null)
@@ -230,10 +257,17 @@ public class IngestionService
             {
                 Name = facilityName, Type = FacilityType.Clinic, City = city, DistrictId = district?.Id,
                 Phone = !string.IsNullOrWhiteSpace(d?.Phone) ? d!.Phone!.Trim() : parsed.Phone, IsVerified = false,
+                Lat = unnamed ? null : raw.Lat, Lng = unnamed ? null : raw.Lng,   // approx. Divar map center
             };
             _db.Facilities.Add(facility);
             facilities.Add(facility);   // so later listings in this run match it too
         }
+        else if (!unnamed && facility.Lat is null && facility.Lng is null && raw.Lat is not null)
+        {
+            // Backfill coords only when the matched (real, named) facility has none — never overwrite a
+            // real (employer-set or verified) location with Divar's fuzzy point.
+            facility.Lat = raw.Lat; facility.Lng = raw.Lng;
+        }
 
         if (kindStr.Contains("job") || kindStr.Contains("استخدام"))
         {
@@ -278,24 +312,33 @@ public class IngestionService
         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>
+    /// <summary>Resolve a role name to an existing Role; if it's genuinely new, create it (dynamic
+    /// taxonomy). Matching is layered so a differently-worded-but-same-meaning role maps to the
+    /// canonical one instead of forking: (1) exact normalized name, (2) synonym/abbreviation alias
+    /// → canonical (دکتر→پزشک عمومی، نرس→پرستار…), (3) create. Only TRUE synonyms collapse — real
+    /// sub-specialties («پرستار ICU») stay distinct on purpose.</summary>
     private Role ResolveOrCreateRole(List<Role> roles, string name, string? category)
     {
         var norm = NormalizeFa(name);
+
+        // (1) Already a known role (same word or spelling variant).
         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));
+        // (2) A synonym of a canonical role → use that role; don't create a duplicate.
+        if (RoleAliases.TryGetValue(norm, out var canonical))
+        {
+            var canonNorm = NormalizeFa(canonical);
+            var aliased = roles.FirstOrDefault(r => NormalizeFa(r.Name) == canonNorm);
+            if (aliased is not null) return aliased;
+            name = canonical; norm = canonNorm;   // canonical not seeded yet → create under its proper name
+        }
 
+        // (3) Genuinely new role — create it under a canonical-resolved category.
         var created = new Role
         {
             Name = Clamp(name.Trim(), 100),                  // respect Role.Name MaxLength(100)
-            Category = Clamp(existingCat ?? wantCat, 50),  // respect Role.Category MaxLength(50)
+            Category = Clamp(ResolveCategory(roles, category), 50),  // respect Role.Category MaxLength(50)
             IsActive = true,
             SortOrder = (roles.Count == 0 ? 0 : roles.Max(r => r.SortOrder)) + 1,
         };
@@ -306,6 +349,58 @@ public class IngestionService
         return created;
     }
 
+    /// <summary>Map an AI-suggested category to a canonical one: synonym alias first
+    /// (پزشکی→پزشک، nursing→پرستار…), then any existing category that normalizes the same, else as-is.</summary>
+    private static string ResolveCategory(List<Role> roles, string? category)
+    {
+        var raw = string.IsNullOrWhiteSpace(category) ? "سایر" : category!.Trim();
+        // Resolve to a canonical first (synonym alias), then to whichever normalized form is the
+        // matching target. Crucially, ALWAYS prefer a category string already stored on a role — even
+        // after an alias maps to a canonical — so we never fork a second variant of the same group.
+        var target = CategoryAliases.TryGetValue(NormalizeFa(raw), out var canonical) ? canonical : raw;
+        var targetNorm = NormalizeFa(target);
+        return roles.Select(r => r.Category)
+            .FirstOrDefault(c => !string.IsNullOrWhiteSpace(c) && NormalizeFa(c) == targetNorm) ?? target;
+    }
+
+    // Synonyms/abbreviations → canonical ROLE name, so the AI naming a role differently maps onto an
+    // existing role instead of forking the taxonomy. Keys are matched after NormalizeFa. Add freely.
+    private static readonly Dictionary<string, string> RoleAliases = BuildAliasMap(new()
+    {
+        ["پزشک عمومی"]             = new[] { "دکتر", "طبیب", "پزشک", "جی پی", "gp", "general practitioner" },
+        ["پزشک متخصص"]            = new[] { "متخصص", "فوق تخصص", "اسپشالیست", "specialist" },
+        ["پرستار"]                 = new[] { "نرس", "nurse", "پرستاری", "کارشناس پرستاری" },
+        ["پرستار سالمندان"]        = new[] { "مراقب سالمند", "مراقب سالمندان", "پرستار سالمند", "نگهدار سالمند", "مراقبت سالمند" },
+        ["ماما"]                   = new[] { "مامایی", "کارشناس مامایی", "midwife" },
+        ["تکنسین اتاق عمل"]        = new[] { "اتاق عمل", "اسکراب", "scrub", "تکنولوژیست اتاق عمل" },
+        ["تکنسین فوریت‌های پزشکی"]  = new[] { "فوریت پزشکی", "تکنسین اورژانس", "پارامدیک", "paramedic", "emt", "اورژانس ۱۱۵" },
+        ["کارشناس آزمایشگاه"]      = new[] { "علوم آزمایشگاهی", "تکنسین آزمایشگاه", "آزمایشگاهی", "لابراتوار", "lab", "laboratory" },
+        ["دندانپزشک"]             = new[] { "دندان پزشک", "دندون پزشک", "dentist" },
+    });
+
+    // Synonyms → canonical CATEGORY (the role-group used for filters/chips).
+    private static readonly Dictionary<string, string> CategoryAliases = BuildAliasMap(new()
+    {
+        ["پزشک"]      = new[] { "دکتر", "طبیب", "doctor", "پزشکی" },
+        ["پرستار"]    = new[] { "پرستاری", "nurse", "nursing" },
+        ["ماما"]      = new[] { "مامایی", "midwifery" },
+        ["تکنسین"]    = new[] { "تکنیسین", "تکنولوژیست", "technician", "کاردان فنی" },
+        ["دندانپزشک"] = new[] { "دندان پزشک", "دندانپزشکی", "dental" },
+    });
+
+    /// <summary>Flatten {canonical → [synonyms]} into a {normalized synonym → canonical} lookup,
+    /// also mapping each canonical's own normalized form to itself.</summary>
+    private static Dictionary<string, string> BuildAliasMap(Dictionary<string, string[]> src)
+    {
+        var map = new Dictionary<string, string>();
+        foreach (var (canonical, aliases) in src)
+        {
+            map[NormalizeFa(canonical)] = canonical;
+            foreach (var a in aliases) map[NormalizeFa(a)] = canonical;
+        }
+        return map;
+    }
+
     /// <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(

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

@@ -3,7 +3,7 @@ using JobsMedical.Web.Models;
 
 namespace JobsMedical.Web.Services.Scraping;
 
-public record ValidationResult(bool IsValid, bool IsSpam, int Confidence, List<string> Issues);
+public record ValidationResult(bool IsValid, bool IsSpam, int Confidence, List<string> Issues, bool LooksMedical = false);
 
 /// <summary>
 /// Scores a parsed listing for completeness and screens out spam. A listing must look like a
@@ -64,7 +64,7 @@ public class ListingValidator
         if (isPromo)
         {
             issues.Add("آگهی تبلیغاتی/آموزشی است، نه استخدام/شیفت");
-            return new ValidationResult(false, true, 0, issues);   // IsSpam → auto-discard
+            return new ValidationResult(false, true, 0, issues, looksMedical);   // IsSpam → auto-discard
         }
 
         // «آماده به کار»: a worker offering themselves. No facility/shift-date expected; the role
@@ -84,7 +84,7 @@ public class ListingValidator
             if (tlen < 20) { ts -= 20; issues.Add("متن خیلی کوتاه است"); }
             ts = Math.Clamp(ts, 0, 100);
             bool tValid = !isSpam && looksMedical && ts >= 50;   // role(40)+medical(10) passes w/o phone
-            return new ValidationResult(tValid, isSpam, ts, issues);
+            return new ValidationResult(tValid, isSpam, ts, issues, looksMedical);
         }
 
         int score = 0;
@@ -107,6 +107,6 @@ public class ListingValidator
 
         // Valid enough for the queue if it's medical, not spam, and reasonably complete.
         bool isValid = !isSpam && looksMedical && score >= 50;
-        return new ValidationResult(isValid, isSpam, score, issues);
+        return new ValidationResult(isValid, isSpam, score, issues, looksMedical);
     }
 }