Fixes from the end-to-end pass, and compose points at rc2

.github/workflows/release.yml

@@ -9,7 +9,7 @@ on:
   workflow_dispatch:
     inputs:
       tag:
-        description: "Image tag to publish (e.g. 0.1.0-rc1)"
+        description: "Image tag to publish (e.g. 0.1.0-rc2)"
         required: true
 
 permissions:

crates/utopia-core/src/models.rs

@@ -194,6 +194,9 @@ pub struct AuditEventView {
     pub target_kind: String,
     pub target_id: Option<Uuid>,
     pub detail: serde_json::Value,
+    /// NULL = 引擎自动（裁决器合并、一致性检查、推理物化……）。界面靠它把
+    /// 「没有人」和「人已被移除」分开：后者 actor_id 还在，只是查不到显示名
+    pub actor_id: Option<Uuid>,
     pub actor_name: Option<String>,
     pub created_at: DateTime<Utc>,
 }

crates/utopia-server/src/api/kbs.rs

@@ -359,14 +359,30 @@ async fn install_packs(
     if pack_ids.is_empty() {
         return Ok(());
     }
+    let mut packs = Vec::with_capacity(pack_ids.len());
     for id in pack_ids {
         let pack = crate::ontology_packs::get(id)
             .ok_or_else(|| AppError::invalid("unknown_pack", format!("未知的本体包：{id}")))?;
-        let bytes = crate::ontology_packs::bytes(pack)?;
-        crate::owl_import::apply(state, kb_id, actor, pack.filename, &bytes)
+        packs.push((pack, crate::ontology_packs::bytes(pack)?));
+    }
+    for (pack, bytes) in &packs {
+        crate::owl_import::apply(state, kb_id, actor, pack.filename, bytes)
             .await
             .map_err(|e| AppError::Other(anyhow::anyhow!("装本体包 {} 失败：{e}", pack.id)))?;
     }
+    // 第二遍：跨包的 domain / range。包是挨个装的，先装的看不见后装的类——
+    // W3C Org 的 headOf 要等 FOAF 的 Agent（#222）。只装一个包时没有"别的包"
+    if packs.len() > 1 {
+        for (pack, bytes) in &packs {
+            let (d, r) =
+                crate::owl_import::relink_domains_ranges(state, kb_id, pack.filename, bytes)
+                    .await
+                    .map_err(|e| {
+                        AppError::Other(anyhow::anyhow!("补本体包 {} 的签名失败：{e}", pack.id))
+                    })?;
+            tracing::debug!(%kb_id, pack = pack.id, domains = d, ranges = r, "跨包签名补链");
+        }
+    }
     Ok(())
 }
 

crates/utopia-server/src/api/sources_routes.rs

@@ -285,6 +285,10 @@ pub async fn sync_now(
 #[derive(Deserialize)]
 pub struct IngestBody {
     pub filename: String,
+    /// 墓碑推送（`deleted: true`）可以不带 content——指南一直这么写，而字段
+    /// 从前是必填，缺了就在反序列化那一步被拒。其余情况仍然必填：空串在
+    /// 下面的校验里挡
+    #[serde(default)]
     pub content: String,
     #[serde(default)]
     pub doc_time: Option<DateTime<Utc>>,
@@ -341,14 +345,33 @@ pub async fn ingest(
     Ok(Json(json!({ "action": action_str(action) })))
 }
 
+/// 推送失败的两种性质。**调用方发错了**和**我们这边没接住**得分开：前者记进
+/// run 供集成调试，但不算来源同步失败——来源没坏，是那一次请求不合格；
+/// 后者才该把来源标成 failed 并进告警中心。从前两者都走 `finish_sync(error)`，
+/// 一次格式错误就让铃铛说"来源同步失败，没有新内容进来"
+enum PushError {
+    /// 4xx：负载不合格（JSON 解析、缺字段）
+    Rejected(String),
+    /// 摄入本身失败
+    Failed(String),
+}
+
+impl PushError {
+    fn message(&self) -> &str {
+        match self {
+            PushError::Rejected(m) | PushError::Failed(m) => m,
+        }
+    }
+}
+
 /// 认证之后的推送处理：解析 + 校验 + 摄入/墓碑。错误一律返回文字（记进 run）。
 async fn handle_push(
     state: &AppState,
     source: &utopia_core::models::Source,
     bytes: &[u8],
-) -> Result<crate::ingest_sources::IngestAction, String> {
-    let body: IngestBody =
-        serde_json::from_slice(bytes).map_err(|e| format!("Invalid JSON payload: {e}"))?;
+) -> Result<crate::ingest_sources::IngestAction, PushError> {
+    let body: IngestBody = serde_json::from_slice(bytes)
+        .map_err(|e| PushError::Rejected(format!("Invalid JSON payload: {e}")))?;
     let identity = body
         .external_id
         .as_deref()
@@ -357,20 +380,24 @@ async fn handle_push(
         .unwrap_or_else(|| body.filename.trim())
         .to_string();
     if identity.is_empty() {
-        return Err("external_id or filename is required".into());
+        return Err(PushError::Rejected(
+            "external_id or filename is required".into(),
+        ));
     }
     let key = format!("api:{identity}");
 
     // 墓碑：标记"不在来源中"（与 custom 的 deleted[] 同一条路径），content 可省略
     if body.deleted {
         utopia_store::documents::mark_missing_keys(&state.pool, source.id, &[key])
             .await
-            .map_err(|e| e.to_string())?;
+            .map_err(|e| PushError::Failed(e.to_string()))?;
         return Ok(crate::ingest_sources::IngestAction::Tombstoned);
     }
 
     if body.filename.trim().is_empty() || body.content.trim().is_empty() {
-        return Err("filename and content are required".into());
+        return Err(PushError::Rejected(
+            "filename and content are required".into(),
+        ));
     }
     let action = crate::ingest_sources::ingest_item(
         state,
@@ -383,11 +410,11 @@ async fn handle_push(
         body.doc_time,
     )
     .await
-    .map_err(|e| e.to_string())?;
+    .map_err(|e| PushError::Failed(e.to_string()))?;
     // 失而复得：曾被墓碑标记的身份再次正常推送，摘掉 missing 标记
     utopia_store::documents::clear_missing_keys(&state.pool, source.id, &[key])
         .await
-        .map_err(|e| e.to_string())?;
+        .map_err(|e| PushError::Failed(e.to_string()))?;
     Ok(action)
 }
 
@@ -434,10 +461,14 @@ pub async fn push(
             state.emit_source(source.kb_id);
             Ok(Json(json!({ "action": action_str(action) })))
         }
-        Err(msg) => {
+        Err(err) => {
+            let msg = err.message().to_string();
             utopia_store::sources::finish_run(&state.pool, run, source.id, Some(&msg), 0, 0)
                 .await?;
-            utopia_store::sources::finish_sync(&state.pool, source.id, Some(&msg), 0).await?;
+            // 只有我们这边没接住才算来源失败；调用方发错了留在 run 历史里就够
+            if let PushError::Failed(_) = err {
+                utopia_store::sources::finish_sync(&state.pool, source.id, Some(&msg), 0).await?;
+            }
             state.emit_source(source.kb_id);
             Err(utopia_core::AppError::Validation(msg).into())
         }

crates/utopia-server/src/extraction.rs

@@ -300,6 +300,8 @@ async fn run(state: &AppState, document_id: Uuid, proposed_by: Option<Uuid>) ->
     utopia_store::documents::set_graph_status(&state.pool, document_id, "extracting").await?;
     state.emit_document(doc.kb_id, document_id);
     let etypes = utopia_store::graph::entity_types(&state.pool, doc.kb_id).await?;
+    // 这一轮落过的事实（新建或重复观察）：结尾对它们跑一遍签名检查
+    let mut touched_facts: Vec<Uuid> = Vec::new();
     let rtypes = utopia_store::graph::relation_types(&state.pool, doc.kb_id).await?;
     // 关系与属性分道：属性走字面值通道，不进关系清单。
     //
@@ -762,6 +764,7 @@ async fn run(state: &AppState, document_id: Uuid, proposed_by: Option<Uuid>) ->
                     confidence,
                 )
                 .await?;
+                touched_facts.push(fact_id);
                 // 属性谓词也留原词：模型偶尔照抄 "person.salary" 全限定名，
                 // 命中的是剥掉前缀后的 key，原样是什么值得留着
                 utopia_store::graph::add_evidence(
@@ -882,6 +885,7 @@ async fn run(state: &AppState, document_id: Uuid, proposed_by: Option<Uuid>) ->
                     confidence,
                 )
                 .await?;
+                touched_facts.push(fact_id);
                 utopia_store::graph::add_evidence(
                     &state.pool,
                     fact_id,
@@ -1126,6 +1130,7 @@ async fn run(state: &AppState, document_id: Uuid, proposed_by: Option<Uuid>) ->
                     confidence,
                 )
                 .await?;
+                touched_facts.push(fact_id);
                 // 重复观察也要挂证据：多来源相互印证，任一来源删除后事实不孤儿化。
                 // 表层谓词随每次观察落笔——甲块说 "runs on"、乙块说 "optimized for"
                 // 会并进同一条事实，放事实上就是先写者胜，放证据上两个都留着
@@ -1221,6 +1226,36 @@ async fn run(state: &AppState, document_id: Uuid, proposed_by: Option<Uuid>) ->
     if let Some(msg) = incomplete_reason(&unextracted, chunks.len()) {
         return Err(anyhow::anyhow!(msg));
     }
+    // 刚落的事实立刻过一遍签名。写入时只掰方向（judge_direction）；掰不动的
+    // ——两个方向都对不上、或宾语没类型判不了——从前要等人按 Review 里的
+    // Run check 才露面，Axioms 一直是 0，图里却躺着反向事实（#222）。
+    // 检查失败不影响抽取本身：事实已经在库里，下一次 Run check 仍然查得到
+    if !touched_facts.is_empty() {
+        match utopia_store::reasoning::signature_breaks(
+            &state.pool,
+            doc.kb_id,
+            Some(&touched_facts),
+        )
+        .await
+        {
+            Ok(broken) if !broken.is_empty() => {
+                match utopia_store::reasoning::record_signature_breaks(
+                    &state.pool,
+                    doc.kb_id,
+                    &broken,
+                )
+                .await
+                {
+                    Ok(_) => state.emit_review(doc.kb_id),
+                    Err(e) => {
+                        tracing::warn!(%document_id, error = %e, "抽取后的签名违规没记进队列")
+                    }
+                }
+            }
+            Ok(_) => {}
+            Err(e) => tracing::warn!(%document_id, error = %e, "抽取后的签名检查失败"),
+        }
+    }
     utopia_store::documents::set_graph_status(&state.pool, document_id, "done").await?;
     state.emit_document(doc.kb_id, document_id);
 

crates/utopia-server/src/mappings.rs

@@ -15,6 +15,39 @@ use uuid::Uuid;
 
 const MAX_SCHEMA_CHARS: usize = 12_000;
 
+/// 探索把 schema 里的量与维度落成 Metric / Dimension 实体，而这两个类不在任何
+/// 内置本体包里——0009 之后建库不再自带类。没有它们，下面的 `type_id` 查不到，
+/// 每条提议都被 `continue` 吞掉，页面只说"已排队"就再无下文（#223）。
+/// 所以探索前把两个类补上：builtin，描述给抽取提示词，本体页可以改
+async fn ensure_concept_types(pool: &sqlx::PgPool, kb_id: Uuid) -> anyhow::Result<()> {
+    for (key, label, description) in [
+        (
+            "metric",
+            "Metric",
+            "An aggregatable business quantity (revenue, order count, average ticket) that              maps to a definition in a mounted database.",
+        ),
+        (
+            "dimension",
+            "Dimension",
+            "A group-by attribute (region, month, product line) that maps to a column in a              mounted database.",
+        ),
+    ] {
+        sqlx::query(
+            "INSERT INTO entity_types (id, kb_id, key, label, builtin, description)
+             SELECT $1, $2, $3, $4, TRUE, $5
+             WHERE NOT EXISTS (SELECT 1 FROM entity_types WHERE kb_id = $2 AND key = $3)",
+        )
+        .bind(Uuid::now_v7())
+        .bind(kb_id)
+        .bind(key)
+        .bind(label)
+        .bind(description)
+        .execute(pool)
+        .await?;
+    }
+    Ok(())
+}
+
 pub async fn explore_mappings(state: &AppState, kb_id: Uuid) -> anyhow::Result<()> {
     let kb = utopia_store::kbs::get(&state.pool, kb_id).await?;
     let settings = utopia_store::settings::get(&state.pool, kb.workspace_id)
@@ -27,6 +60,7 @@ pub async fn explore_mappings(state: &AppState, kb_id: Uuid) -> anyhow::Result<(
     if sources.is_empty() {
         anyhow::bail!("No data sources mounted");
     }
+    ensure_concept_types(&state.pool, kb_id).await?;
 
     // 各源 schema（引擎直读，保证新鲜；限量防 prompt 爆炸）
     let mut schema_txt = String::new();
@@ -168,6 +202,27 @@ pub async fn explore_mappings(state: &AppState, kb_id: Uuid) -> anyhow::Result<(
     }
 
     tracing::info!(%kb_id, proposals = accepted, "映射探索完成，提议已入审核队列");
+    // 一条都没提出来时页面上什么都不会变——Pending 还是 0，而"已排队"那句
+    // 早就翻篇了。走告警中心说一声，人才知道该去刷新结构或给列加注释
+    if accepted == 0 {
+        if let Err(e) = utopia_store::alerts::raise(
+            &state.pool,
+            utopia_store::alerts::NewAlert {
+                kb_id: Some(kb_id),
+                severity: "info",
+                kind: utopia_store::alerts::kind::MAPPING_EXPLORATION_EMPTY,
+                min_role: utopia_core::models::Role::Editor,
+                subject_type: None,
+                subject_id: None,
+                detail: serde_json::json!({ "proposals": 0, "sources": source_names }),
+            },
+        )
+        .await
+        {
+            tracing::warn!(%kb_id, error = %e, "映射探索空结果的告警没写进去");
+        }
+        state.emit_alert();
+    }
     state.emit_review(kb_id);
     Ok(())
 }

crates/utopia-server/src/owl_import.rs

@@ -323,6 +323,64 @@ pub async fn plan(
     Ok((plan, proj, format))
 }
 
+/// 装完一组包之后再过一遍：把 domain / range 指向**别的包里的类**的那些关系接上。
+///
+/// 单次导入认本文件里的类和库里已有的类（见 [`apply`] 里的 resolve），但包是挨个
+/// 装的：装 W3C Org 时 FOAF 还没来，`headOf` 的 `rdfs:domain foaf:Agent` 就落了空；
+/// 等 FOAF 装好，没人回头补。没有 domain 的谓词 `judge_direction` 不判方向，
+/// 反向的 `Project Aurora head_of Li Ting` 就原样进图（#222）。
+///
+/// 只补不删，关联表 ON CONFLICT DO NOTHING，重复跑无害。属性不在这里：属性的
+/// 去向在计划阶段就定了（没有 domain 的根本建不出来），事后补 domain 改不了它
+/// 已经是不是一列的事实
+pub async fn relink_domains_ranges(
+    state: &AppState,
+    kb_id: Uuid,
+    filename: &str,
+    bytes: &[u8],
+) -> AppResult<(usize, usize)> {
+    let format = RdfFormat::detect(filename, bytes);
+    let proj = ontology_rdf::project(bytes, format).map_err(|e| {
+        utopia_core::AppError::invalid_detail(
+            "bad_ontology_file",
+            "Could not parse this ontology file",
+            e.to_string(),
+        )
+    })?;
+    let classes: HashMap<String, Uuid> = utopia_store::graph::entity_types(&state.pool, kb_id)
+        .await?
+        .into_iter()
+        .filter_map(|t| t.iri.clone().map(|i| (i, t.id)))
+        .collect();
+    let relations: HashMap<String, Uuid> = utopia_store::graph::relation_types(&state.pool, kb_id)
+        .await?
+        .into_iter()
+        .filter_map(|r| r.iri.clone().map(|i| (i, r.id)))
+        .collect();
+    let mut link_d: Vec<(Uuid, Uuid)> = Vec::new();
+    let mut link_r: Vec<(Uuid, Uuid)> = Vec::new();
+    for p in &proj.properties {
+        if p.is_datatype {
+            continue;
+        }
+        let Some(&rid) = relations.get(&p.iri) else {
+            continue;
+        };
+        link_d.extend(
+            p.domains
+                .iter()
+                .filter_map(|d| classes.get(d).map(|&t| (rid, t))),
+        );
+        link_r.extend(
+            p.ranges
+                .iter()
+                .filter_map(|r| classes.get(r).map(|&t| (rid, t))),
+        );
+    }
+    utopia_store::ontology::link_domains_ranges_bulk(&state.pool, &link_d, &link_r).await?;
+    Ok((link_d.len(), link_r.len()))
+}
+
 /// 执行计划。属性在类之后落库——它们要挂在 domain 上，而 domain 要等
 /// 类先建好并解析 IRI → id（就是下面那个 `id_of`）。
 pub async fn apply(
@@ -500,9 +558,20 @@ pub async fn apply(
             continue;
         };
         // domain/range 指向没被建出来的类时只丢那一个，不丢整条关系：
-        // 关系不像属性那样必须挂在类上，没有 domain 就是"不限主语类型"
+        // 关系不像属性那样必须挂在类上，没有 domain 就是"不限主语类型"。
+        //
+        // **库里已有的类也算数。** 从前只认本文件里的类，于是 W3C Org 的
+        // `headOf rdfs:domain foaf:Agent` 在 FOAF 已经装好的库里照样丢 domain，
+        // 而没有 domain 的谓词 `judge_direction` 根本不判方向（#222）
         let resolve = |iris: &[String]| -> Vec<Uuid> {
-            iris.iter().filter_map(|i| id_of.get(i).copied()).collect()
+            iris.iter()
+                .filter_map(|i| {
+                    id_of
+                        .get(i)
+                        .copied()
+                        .or_else(|| existing_by_iri.get(i).copied())
+                })
+                .collect()
         };
         let domains = resolve(&p.domains);
         let ranges = resolve(&p.ranges);