feat(ingest): 진행 로그 개선 — 파일명/phase/heartbeat/slowest 요약

OCR/caption 켜진 볼트 ingest 가 중간부터 느릴 때 TTY 진행바가 파일명·phase·
모델·경과시간을 안 보여 "멈춤"처럼 보이던 문제 해결.
- 신규 wire AssetPhase{idx,total,phase,model} + AssetTimings.ocr_ms/caption_ms
  (additive, ingest_progress.v1 유지)
- app: apply_ocr/apply_caption/embed 진입 시 AssetPhase emit + ocr/caption 시간 측정
- cli: TTY 진행바에 현재 파일명 + phase(model) + asset 경과초(heartbeat),
  종료 시 최장 소요 파일 top-5 요약(quiet 여도 출력, --json 미출력)
- wire schema / README / HANDOFF / HOTFIXES 동기화, version 0.26.0 → 0.27.0

검증(리더): clippy 0, kebab-app/cli 61그룹·parse-image/tui 14그룹 0실패(-j8).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

crates/kebab-app/src/ingest_progress.rs

@@ -96,14 +96,33 @@ pub enum IngestEvent {
     /// `idx/total` while its per-chunk phases churn. `chunks` is the chunk
     /// count for asset `idx`.
     AssetChunked { idx: u32, total: u32, chunks: u32 },
+    /// v0.26.0 (additive): emitted when an asset enters a *slow* internal
+    /// phase, so the interactive progress bar can show **which** phase
+    /// (and which model) is currently running instead of looking frozen.
+    /// `phase` ∈ {`"ocr"`, `"caption"`, `"embed"`}; short phases
+    /// (parse / chunk / store) are intentionally *not* emitted to avoid
+    /// noise. `model` is the model performing the phase — the vision LLM
+    /// id for `ocr` / `caption`, the embedder `model_id` for `embed`
+    /// (`None` when the phase runs without a configured model, e.g. embed
+    /// with no embedder wired). Emitted once per (asset, phase); no
+    /// throttle needed (low frequency). Wire v1 consumers that predate
+    /// this variant simply ignore the unknown `asset_phase` kind.
+    AssetPhase {
+        idx: u32,
+        total: u32,
+        phase: String,
+        model: Option<String>,
+    },
     /// v0.24.0 (additive): per-phase wall-clock (milliseconds) for asset
-    /// `idx`, emitted once the asset's markdown pipeline finishes. Lets a
-    /// user see *where* the time went (parse / chunk / embed / store)
-    /// without parsing logs. Only the markdown path emits this; the
-    /// image / PDF paths surface `AssetChunked` but skip phase timing (their
-    /// phase shapes differ — OCR / caption). `expansion_ms` is retained for
-    /// wire compatibility but is always 0 since doc-side expansion was
-    /// removed (HOTFIXES 2026-06-03).
+    /// `idx`, emitted once the asset's pipeline finishes. Lets a user see
+    /// *where* the time went (parse / chunk / ocr / caption / embed /
+    /// store) without parsing logs. The markdown path leaves `ocr_ms` /
+    /// `caption_ms` at 0 (no image analysis); the image / PDF paths fill
+    /// them so the slowest-asset summary attributes vision-model time
+    /// correctly. `expansion_ms` is retained for wire compatibility but is
+    /// always 0 since doc-side expansion was removed (HOTFIXES 2026-06-03).
+    /// `ocr_ms` / `caption_ms` (v0.26.0) are additive with serde default 0
+    /// so pre-v0.26.0 consumers deserialize cleanly.
     AssetTimings {
         idx: u32,
         total: u32,
@@ -112,6 +131,10 @@ pub enum IngestEvent {
         expansion_ms: u64,
         embed_ms: u64,
         store_ms: u64,
+        #[serde(default)]
+        ocr_ms: u64,
+        #[serde(default)]
+        caption_ms: u64,
     },
     /// Run finished normally. `counts` is the final aggregate.
     Completed { counts: AggregateCounts },
@@ -261,19 +284,23 @@ mod tests {
             expansion_ms: 45_000,
             embed_ms: 800,
             store_ms: 20,
+            ocr_ms: 1_200,
+            caption_ms: 3_400,
         };
         let v = serde_json::to_value(&ev).unwrap();
         assert_eq!(
             v.get("kind").and_then(|s| s.as_str()),
             Some("asset_timings")
         );
-        // All five phase fields are present (plain u64, always serialized).
+        // All phase fields are present (plain u64, always serialized).
         for (field, want) in [
             ("parse_ms", 12u64),
             ("chunk_ms", 3),
             ("expansion_ms", 45_000),
             ("embed_ms", 800),
             ("store_ms", 20),
+            ("ocr_ms", 1_200),
+            ("caption_ms", 3_400),
         ] {
             assert_eq!(
                 v.get(field).and_then(serde_json::Value::as_u64),
@@ -283,6 +310,64 @@ mod tests {
         }
     }
 
+    #[test]
+    fn asset_timings_ocr_caption_default_to_zero_for_legacy_wire() {
+        // v0.26.0 additive: a pre-v0.26.0 wire payload omits ocr_ms /
+        // caption_ms; serde `default` must fill 0 so old producers stay
+        // compatible.
+        let legacy = serde_json::json!({
+            "kind": "asset_timings",
+            "idx": 1, "total": 1,
+            "parse_ms": 5, "chunk_ms": 2, "expansion_ms": 0,
+            "embed_ms": 10, "store_ms": 3
+        });
+        let ev: IngestEvent = serde_json::from_value(legacy).unwrap();
+        match ev {
+            IngestEvent::AssetTimings {
+                ocr_ms,
+                caption_ms,
+                embed_ms,
+                ..
+            } => {
+                assert_eq!(ocr_ms, 0);
+                assert_eq!(caption_ms, 0);
+                assert_eq!(embed_ms, 10);
+            }
+            other => panic!("unexpected event: {other:?}"),
+        }
+    }
+
+    #[test]
+    fn asset_phase_serializes_with_discriminator() {
+        // v0.26.0 additive variant — `kind` must be snake_case
+        // `asset_phase`, `phase` is the slow-phase label, `model` the
+        // model id (nullable).
+        let ev = IngestEvent::AssetPhase {
+            idx: 4,
+            total: 12,
+            phase: "ocr".into(),
+            model: Some("gemma4:e4b".into()),
+        };
+        let v = serde_json::to_value(&ev).unwrap();
+        assert_eq!(v.get("kind").and_then(|s| s.as_str()), Some("asset_phase"));
+        assert_eq!(v.get("idx").and_then(serde_json::Value::as_u64), Some(4));
+        assert_eq!(v.get("phase").and_then(|s| s.as_str()), Some("ocr"));
+        assert_eq!(v.get("model").and_then(|s| s.as_str()), Some("gemma4:e4b"));
+    }
+
+    #[test]
+    fn asset_phase_model_none_serializes_as_null() {
+        let ev = IngestEvent::AssetPhase {
+            idx: 1,
+            total: 1,
+            phase: "embed".into(),
+            model: None,
+        };
+        let v = serde_json::to_value(&ev).unwrap();
+        assert_eq!(v.get("phase").and_then(|s| s.as_str()), Some("embed"));
+        assert!(v.get("model").is_some_and(serde_json::Value::is_null));
+    }
+
     #[test]
     fn ingest_event_completed_has_counts() {
         let ev = IngestEvent::Completed {

crates/kebab-app/src/lib.rs

@@ -1350,6 +1350,17 @@ fn ingest_one_asset(
     let store_ms = u64::try_from(t_store.elapsed().as_millis()).unwrap_or(u64::MAX);
 
     // Embed + vector upsert (only when both sides are configured).
+    // v0.26.0: surface the embed phase + model so a long embed run reads as
+    // "embedding(<model>)…" rather than a frozen bar (markdown path too).
+    crate::ingest_progress::emit(
+        progress,
+        crate::ingest_progress::IngestEvent::AssetPhase {
+            idx,
+            total,
+            phase: "embed".to_string(),
+            model: embedder.map(|e| e.model_id().0),
+        },
+    );
     let t_embed = std::time::Instant::now();
     // Stale-vector purge is LanceDB I/O, so it belongs to the embed/vector
     // phase — not the SQLite `store` phase. Keeping it here makes `store_ms`
@@ -1414,7 +1425,8 @@ fn ingest_one_asset(
 
     let embed_ms = u64::try_from(t_embed.elapsed().as_millis()).unwrap_or(u64::MAX);
 
-    // v0.24.0: phase-timing breakdown for this asset (markdown path only).
+    // v0.24.0: phase-timing breakdown for this asset (markdown path).
+    // ocr_ms / caption_ms are 0 — markdown has no image-analysis phases.
     crate::ingest_progress::emit(
         progress,
         crate::ingest_progress::IngestEvent::AssetTimings {
@@ -1425,6 +1437,8 @@ fn ingest_one_asset(
             expansion_ms,
             embed_ms,
             store_ms,
+            ocr_ms: 0,
+            caption_ms: 0,
         },
     );
 
@@ -1545,9 +1559,11 @@ fn ingest_one_image_asset(
         workspace_root: &workspace_root,
         config: &extract_config,
     };
+    let t_parse = std::time::Instant::now();
     let mut canonical = app
         .extract_for(&asset.media_type, &ctx, &bytes)
         .context("kb-app::extract_for (image)")?;
+    let parse_ms = u64::try_from(t_parse.elapsed().as_millis()).unwrap_or(u64::MAX);
 
     // 2 + 3. Apply OCR / caption when their adapters exist. Both are
     //        Lenient — failure is captured into Provenance Warning,
@@ -1562,44 +1578,74 @@ fn ingest_one_image_asset(
     let lang_hint = lang_hint_from_doc(&canonical);
     let now = time::OffsetDateTime::now_utc();
     let mut warning_notes: Vec<String> = Vec::new();
+    // v0.26.0: vision phases (OCR / caption) are the usual bottleneck on an
+    // image-heavy vault and emitted no progress before — so the bar looked
+    // frozen. Surface each as an `AssetPhase` and measure its wall-clock for
+    // the slowest-asset summary.
+    let mut ocr_ms = 0_u64;
+    let mut caption_ms = 0_u64;
     match canonical.blocks.first_mut() {
         Some(Block::ImageRef(block)) => {
-            if let Some(engine) = ocr_engine
-                && let Err(e) = apply_ocr(
+            if let Some(engine) = ocr_engine {
+                crate::ingest_progress::emit(
+                    progress,
+                    crate::ingest_progress::IngestEvent::AssetPhase {
+                        idx,
+                        total,
+                        phase: "ocr".to_string(),
+                        model: Some(engine.model().to_string()),
+                    },
+                );
+                let t_ocr = std::time::Instant::now();
+                let res = apply_ocr(
                     engine,
                     &bytes,
                     block,
                     lang_hint.as_ref(),
                     &mut canonical.provenance.events,
-                )
-            {
-                record_image_analysis_failure(
-                    asset,
-                    &mut canonical.provenance.events,
-                    &mut warning_notes,
-                    "OcrFailed",
-                    e,
-                    now,
                 );
+                ocr_ms = u64::try_from(t_ocr.elapsed().as_millis()).unwrap_or(u64::MAX);
+                if let Err(e) = res {
+                    record_image_analysis_failure(
+                        asset,
+                        &mut canonical.provenance.events,
+                        &mut warning_notes,
+                        "OcrFailed",
+                        e,
+                        now,
+                    );
+                }
             }
-            if let Some(llm) = caption_llm
-                && let Err(e) = apply_caption(
+            if let Some(llm) = caption_llm {
+                crate::ingest_progress::emit(
+                    progress,
+                    crate::ingest_progress::IngestEvent::AssetPhase {
+                        idx,
+                        total,
+                        phase: "caption".to_string(),
+                        model: Some(llm.model_ref().id),
+                    },
+                );
+                let t_caption = std::time::Instant::now();
+                let res = apply_caption(
                     llm,
                     &bytes,
                     block,
                     lang_hint.as_ref(),
                     &app.config,
                     &mut canonical.provenance.events,
-                )
-            {
-                record_image_analysis_failure(
-                    asset,
-                    &mut canonical.provenance.events,
-                    &mut warning_notes,
-                    "CaptionFailed",
-                    e,
-                    now,
                 );
+                caption_ms = u64::try_from(t_caption.elapsed().as_millis()).unwrap_or(u64::MAX);
+                if let Err(e) = res {
+                    record_image_analysis_failure(
+                        asset,
+                        &mut canonical.provenance.events,
+                        &mut warning_notes,
+                        "CaptionFailed",
+                        e,
+                        now,
+                    );
+                }
             }
         }
         // P6-1 contract: image documents always have exactly one
@@ -1634,12 +1680,13 @@ fn ingest_one_image_asset(
     //    `Block::ImageRef` arm already produces a single chunk per
     //    image (P1-5). The chunk text now follows the (β) plain-concat
     //    contract per the kebab-chunk render_block_text update.
+    let t_chunk = std::time::Instant::now();
     let chunks = MdHeadingV1Chunker
         .chunk(&canonical, chunk_policy)
         .context("kb-chunk::MdHeadingV1Chunker::chunk (image)")?;
+    let chunk_ms = u64::try_from(t_chunk.elapsed().as_millis()).unwrap_or(u64::MAX);
 
-    // v0.24.0: surface chunk count for the image path too (phase timing is
-    // markdown-only, but AssetChunked is consistent across media).
+    // v0.24.0: surface chunk count for the image path too.
     crate::ingest_progress::emit(
         progress,
         crate::ingest_progress::IngestEvent::AssetChunked {
@@ -1656,6 +1703,7 @@ fn ingest_one_image_asset(
     if let Some(emb) = embedder {
         canonical.last_embedding_version = Some(emb.model_version());
     }
+    let t_store = std::time::Instant::now();
     purge_vector_orphans_for_workspace_path(app, asset, vector_store)?;
     app.sqlite
         .put_asset_with_bytes(asset, &bytes)
@@ -1669,7 +1717,18 @@ fn ingest_one_image_asset(
     app.sqlite
         .put_chunks(&canonical.doc_id, &chunks)
         .context("DocumentStore::put_chunks (image)")?;
+    let store_ms = u64::try_from(t_store.elapsed().as_millis()).unwrap_or(u64::MAX);
 
+    crate::ingest_progress::emit(
+        progress,
+        crate::ingest_progress::IngestEvent::AssetPhase {
+            idx,
+            total,
+            phase: "embed".to_string(),
+            model: embedder.map(|e| e.model_id().0),
+        },
+    );
+    let t_embed = std::time::Instant::now();
     if let (Some(emb), Some(vec_store)) = (embedder, vector_store)
         && !chunks.is_empty()
     {
@@ -1710,6 +1769,25 @@ fn ingest_one_image_asset(
             .upsert(&records)
             .context("VectorStore::upsert (image)")?;
     }
+    let embed_ms = u64::try_from(t_embed.elapsed().as_millis()).unwrap_or(u64::MAX);
+
+    // v0.26.0: per-phase timing for the image path — ocr_ms / caption_ms
+    // carry the vision-model cost so the slowest-asset summary attributes
+    // an image-heavy run's bottleneck correctly.
+    crate::ingest_progress::emit(
+        progress,
+        crate::ingest_progress::IngestEvent::AssetTimings {
+            idx,
+            total,
+            parse_ms,
+            chunk_ms,
+            expansion_ms: 0,
+            embed_ms,
+            store_ms,
+            ocr_ms,
+            caption_ms,
+        },
+    );
 
     let kind = if existing_doc_ids.contains(&canonical.doc_id.0) {
         kebab_core::IngestItemKind::Updated
@@ -2053,9 +2131,11 @@ fn ingest_one_pdf_asset(
         workspace_root: &workspace_root,
         config: &extract_config,
     };
+    let t_parse = std::time::Instant::now();
     let mut canonical = app
         .extract_for(&asset.media_type, &ctx, &bytes)
         .context("kb-app::extract_for (pdf)")?;
+    let parse_ms = u64::try_from(t_parse.elapsed().as_millis()).unwrap_or(u64::MAX);
 
     // v0.20 sub-item 1: post-extract OCR enrichment (PR #187 registry
     // dispatch invariant 보존 — extract_for 가 normal entry).
@@ -2191,9 +2271,11 @@ fn ingest_one_pdf_asset(
     // validates every block carries `SourceSpan::Page`; failure here
     // means the parser drifted from its contract.
     let chunker = PdfPageV1Chunker;
+    let t_chunk = std::time::Instant::now();
     let chunks = chunker
         .chunk(&canonical, chunk_policy)
         .context("kb-chunk::PdfPageV1Chunker::chunk")?;
+    let chunk_ms = u64::try_from(t_chunk.elapsed().as_millis()).unwrap_or(u64::MAX);
 
     // v0.24.0: surface chunk count for the PDF path too.
     crate::ingest_progress::emit(
@@ -2212,6 +2294,7 @@ fn ingest_one_pdf_asset(
         canonical.last_embedding_version = Some(emb.model_version());
     }
 
+    let t_store = std::time::Instant::now();
     purge_vector_orphans_for_workspace_path(app, asset, vector_store)?;
     app.sqlite
         .put_asset_with_bytes(asset, &bytes)
@@ -2225,7 +2308,18 @@ fn ingest_one_pdf_asset(
     app.sqlite
         .put_chunks(&canonical.doc_id, &chunks)
         .context("DocumentStore::put_chunks (pdf)")?;
+    let store_ms = u64::try_from(t_store.elapsed().as_millis()).unwrap_or(u64::MAX);
 
+    crate::ingest_progress::emit(
+        progress,
+        crate::ingest_progress::IngestEvent::AssetPhase {
+            idx,
+            total,
+            phase: "embed".to_string(),
+            model: embedder.map(|e| e.model_id().0),
+        },
+    );
+    let t_embed = std::time::Instant::now();
     if let (Some(emb), Some(vec_store)) = (embedder, vector_store)
         && !chunks.is_empty()
     {
@@ -2264,6 +2358,25 @@ fn ingest_one_pdf_asset(
             .upsert(&records)
             .context("VectorStore::upsert (pdf)")?;
     }
+    let embed_ms = u64::try_from(t_embed.elapsed().as_millis()).unwrap_or(u64::MAX);
+
+    // v0.26.0: per-phase timing for the PDF path. `ocr_ms` reuses the
+    // page-OCR total already computed above so a scanned-PDF run's OCR cost
+    // shows up in the slowest-asset summary; caption is markdown/image-only.
+    crate::ingest_progress::emit(
+        progress,
+        crate::ingest_progress::IngestEvent::AssetTimings {
+            idx,
+            total,
+            parse_ms,
+            chunk_ms,
+            expansion_ms: 0,
+            embed_ms,
+            store_ms,
+            ocr_ms: pdf_ocr_ms_total.unwrap_or(0),
+            caption_ms: 0,
+        },
+    );
 
     let kind = if existing_doc_ids.contains(&canonical.doc_id.0) {
         kebab_core::IngestItemKind::Updated

crates/kebab-cli/src/progress.rs

@@ -19,16 +19,23 @@
 //! `Sender` end is dropped (i.e. when `ingest_with_config_progress`
 //! returns).
 
+use std::collections::HashMap;
 use std::io::{IsTerminal, Write};
 use std::sync::mpsc::Receiver;
+use std::sync::{Arc, Mutex};
+use std::time::Instant;
 
-use indicatif::{ProgressBar, ProgressDrawTarget, ProgressStyle};
+use indicatif::{ProgressBar, ProgressDrawTarget, ProgressState, ProgressStyle};
 use kebab_app::IngestEvent;
 use time::OffsetDateTime;
 use time::format_description::well_known::Rfc3339;
 
 use crate::wire;
 
+/// v0.26.0: number of slowest assets surfaced in the end-of-run summary.
+/// Constant for now (spec defers the config knob).
+const SLOWEST_TOP_N: usize = 5;
+
 /// Rendering mode for `ProgressDisplay`. The mode is fixed at
 /// construction — each `kebab ingest` invocation is a single mode
 /// (chosen from `--json` plus `IsTerminal` detection).
@@ -65,11 +72,33 @@ impl ProgressMode {
 pub struct ProgressDisplay {
     mode: ProgressMode,
     bar: Option<ProgressBar>,
+    /// v0.26.0 heartbeat: start `Instant` of the asset currently in
+    /// flight, shared with the bar's steady-tick custom template key so
+    /// the `(Ns)` elapsed counter advances *between* events (the drain
+    /// loop blocks on `recv()`, so without the ticker the counter would
+    /// freeze). `None` while scanning / between assets / after completion.
+    asset_start: Arc<Mutex<Option<Instant>>>,
+    /// v0.26.0: workspace path of the asset currently in flight — set on
+    /// `AssetStarted`, reused by `AssetPhase` to render `{path} · {phase}…`.
+    current_path: Option<String>,
+    /// v0.26.0 slowest summary: idx → path, captured from `AssetStarted`
+    /// so `AssetTimings` (which only carries `idx`) can name the asset.
+    asset_paths: HashMap<u32, String>,
+    /// v0.26.0 slowest summary: (path, total_ms) per asset that reported
+    /// `AssetTimings`. Sorted + truncated to top-N on `Completed`.
+    timings: Vec<(String, u64)>,
 }
 
 impl ProgressDisplay {
     pub fn new(mode: ProgressMode) -> Self {
-        Self { mode, bar: None }
+        Self {
+            mode,
+            bar: None,
+            asset_start: Arc::new(Mutex::new(None)),
+            current_path: None,
+            asset_paths: HashMap::new(),
+            timings: Vec::new(),
+        }
     }
 
     /// Block until `rx` returns `Err` (sender dropped). Renders one
@@ -120,15 +149,43 @@ impl ProgressDisplay {
             }
             IngestEvent::ScanCompleted { total } => {
                 if let Some(bar) = self.bar.as_mut() {
-                    bar.disable_steady_tick();
                     bar.set_length(u64::from(*total));
                     bar.set_position(0);
+                    // v0.26.0: a custom `{asset_elapsed}` key reads the shared
+                    // per-asset start `Instant` and appends ` (Ns)`. Combined
+                    // with the steady tick below, the elapsed counter advances
+                    // even while the drain loop is blocked on `recv()` waiting
+                    // for the next (possibly very slow) phase event.
+                    let asset_start = Arc::clone(&self.asset_start);
                     bar.set_style(
-                        ProgressStyle::with_template("ingest [{bar:30}] {pos}/{len} {wide_msg}")
-                            .unwrap()
-                            .progress_chars("=> "),
+                        ProgressStyle::with_template(
+                            "ingest [{bar:30}] {pos}/{len} {wide_msg}{asset_elapsed}",
+                        )
+                        .unwrap()
+                        .with_key(
+                            "asset_elapsed",
+                            move |_: &ProgressState, w: &mut dyn std::fmt::Write| {
+                                if let Ok(guard) = asset_start.lock()
+                                    && let Some(started) = *guard
+                                {
+                                    let secs = started.elapsed().as_secs();
+                                    // Only show once the asset has been running
+                                    // a moment — avoids `(0s)` flicker on fast
+                                    // assets.
+                                    if secs >= 1 {
+                                        let _ = write!(w, " ({secs}s)");
+                                    }
+                                }
+                            },
+                        )
+                        .progress_chars("=> "),
                     );
                     bar.set_message("");
+                    if tty && !quiet {
+                        bar.enable_steady_tick(std::time::Duration::from_secs(1));
+                    } else {
+                        bar.disable_steady_tick();
+                    }
                 }
                 if !tty && !quiet {
                     let mut err = std::io::stderr().lock();
@@ -141,11 +198,22 @@ impl ProgressDisplay {
                 path,
                 media,
             } => {
+                // v0.26.0: remember the path so AssetPhase can render it and
+                // the slowest summary (keyed by idx in AssetTimings) can name
+                // the asset.
+                self.current_path = Some(path.clone());
+                self.asset_paths.insert(*idx, path.clone());
+                // v0.26.0: (re)start the per-asset heartbeat clock.
+                if let Ok(mut guard) = self.asset_start.lock() {
+                    *guard = Some(Instant::now());
+                }
                 if let Some(bar) = self.bar.as_ref() {
-                    // One draw per file: position only. set_message() would
-                    // trigger a second independent draw and pollute TTY scrollback.
-                    // Filename is visible in the non-TTY plain-line path below.
                     bar.set_position(u64::from(idx.saturating_sub(1)));
+                    // v0.26.0: show the current filename on the bar (TTY).
+                    // Previously position-only — the interactive user couldn't
+                    // tell which file was in flight. The steady tick redraws
+                    // in place, so this no longer pollutes scrollback.
+                    bar.set_message(abbreviate_path(path));
                 }
                 if !tty && !quiet {
                     let mut err = std::io::stderr().lock();
@@ -154,8 +222,35 @@ impl ProgressDisplay {
             }
             IngestEvent::AssetFinished { .. } => {
                 // Position is advanced in AssetStarted; bar.finish_and_clear()
-                // in Completed handles the final state. No per-asset bar update
-                // here avoids the duplicate-frame artifact in TTY scrollback.
+                // in Completed handles the final state. v0.26.0: stop the
+                // heartbeat clock so the bar doesn't show a stale `(Ns)` in the
+                // gap before the next AssetStarted.
+                if let Ok(mut guard) = self.asset_start.lock() {
+                    *guard = None;
+                }
+                self.current_path = None;
+            }
+            // v0.26.0: an asset entered a slow internal phase (ocr / caption /
+            // embed). Surface which phase + model is running so a multi-second
+            // vision-model call no longer looks frozen.
+            IngestEvent::AssetPhase {
+                idx,
+                total,
+                phase,
+                model,
+            } => {
+                let label = match model {
+                    Some(m) => format!("{phase}({m})"),
+                    None => phase.clone(),
+                };
+                if let Some(bar) = self.bar.as_ref() {
+                    let path = self.current_path.as_deref().unwrap_or("");
+                    bar.set_message(format!("{} · {label}…", abbreviate_path(path)));
+                }
+                if !tty && !quiet {
+                    let mut err = std::io::stderr().lock();
+                    let _ = writeln!(err, "ingest: {idx}/{total} · {label}…");
+                }
             }
             // v0.24.0: asset-internal phase visibility. AssetChunked uses the
             // bar *message* (live sub-progress for the current asset) —
@@ -172,31 +267,50 @@ impl ProgressDisplay {
                 }
             }
             IngestEvent::AssetTimings {
+                idx,
                 parse_ms,
                 chunk_ms,
                 embed_ms,
                 store_ms,
+                ocr_ms,
+                caption_ms,
                 ..
             } => {
+                // v0.26.0: accumulate (path, total_ms) for the slowest summary.
+                // total = every measured phase (expansion_ms is always 0).
+                let total_ms = parse_ms + chunk_ms + embed_ms + store_ms + ocr_ms + caption_ms;
+                if let Some(path) = self.asset_paths.get(idx) {
+                    self.timings.push((path.clone(), total_ms));
+                }
                 if let Some(bar) = self.bar.as_ref() {
                     bar.set_message("");
                 }
                 if !quiet {
                     let mut err = std::io::stderr().lock();
-                    let _ = writeln!(
-                        err,
-                        "  ⏱ parse {} · chunk {} · embed {} · store {}",
-                        fmt_ms(*parse_ms),
-                        fmt_ms(*chunk_ms),
-                        fmt_ms(*embed_ms),
-                        fmt_ms(*store_ms),
-                    );
+                    // v0.26.0: only print ocr / caption when they actually ran
+                    // (markdown leaves them 0) so the text path stays uncluttered.
+                    let mut parts = vec![
+                        format!("parse {}", fmt_ms(*parse_ms)),
+                        format!("chunk {}", fmt_ms(*chunk_ms)),
+                    ];
+                    if *ocr_ms > 0 {
+                        parts.push(format!("ocr {}", fmt_ms(*ocr_ms)));
+                    }
+                    if *caption_ms > 0 {
+                        parts.push(format!("caption {}", fmt_ms(*caption_ms)));
+                    }
+                    parts.push(format!("embed {}", fmt_ms(*embed_ms)));
+                    parts.push(format!("store {}", fmt_ms(*store_ms)));
+                    let _ = writeln!(err, "  ⏱ {}", parts.join(" · "));
                 }
             }
             IngestEvent::Completed { counts } => {
                 if let Some(bar) = self.bar.take() {
                     bar.finish_and_clear();
                 }
+                if let Ok(mut guard) = self.asset_start.lock() {
+                    *guard = None;
+                }
                 // Always emit summary in both TTY and non-TTY (unless quiet).
                 // Bug fix: previously TTY had no summary line after bar.finish_and_clear().
                 if !quiet {
@@ -206,6 +320,10 @@ impl ProgressDisplay {
                         "ingest: complete (scanned={} new={} updated={} skipped={} errors={})",
                         counts.scanned, counts.new, counts.updated, counts.skipped, counts.errors,
                     );
+                    // v0.26.0: slowest-asset summary. Useful in both TTY and
+                    // non-TTY (it pinpoints the bottleneck file), so it prints
+                    // unless --quiet. --json mode never reaches here (emit_json).
+                    let _ = write_slowest_summary(&mut err, &self.timings, SLOWEST_TOP_N);
                 }
             }
             IngestEvent::Aborted { counts } => {
@@ -286,6 +404,48 @@ fn fmt_ms(ms: u64) -> String {
     }
 }
 
+/// v0.26.0: shorten an over-long workspace path for the progress-bar
+/// message so the live `(Ns)` heartbeat suffix stays visible on a narrow
+/// terminal. Keeps the tail (filename + a couple of parents) — that's the
+/// distinguishing part — and prefixes `…` when truncated. Paths up to the
+/// budget pass through verbatim.
+fn abbreviate_path(path: &str) -> String {
+    const MAX: usize = 48;
+    let char_count = path.chars().count();
+    if char_count <= MAX {
+        return path.to_string();
+    }
+    // Keep the last MAX-1 chars (1 reserved for the leading ellipsis).
+    let tail: String = path
+        .chars()
+        .skip(char_count - (MAX - 1))
+        .collect::<String>();
+    format!("…{tail}")
+}
+
+/// v0.26.0: render the end-of-run "slowest assets" summary. Sorts
+/// `(path, total_ms)` descending by time, takes the top `n`, and writes a
+/// compact table to `w`. No-op (writes nothing) when `timings` is empty so
+/// a run with no per-asset timing (e.g. all-skipped) prints no stray header.
+fn write_slowest_summary(
+    w: &mut impl Write,
+    timings: &[(String, u64)],
+    n: usize,
+) -> std::io::Result<()> {
+    if timings.is_empty() {
+        return Ok(());
+    }
+    let mut sorted: Vec<&(String, u64)> = timings.iter().collect();
+    // desc by ms; ties broken by path for deterministic output.
+    sorted.sort_by(|a, b| b.1.cmp(&a.1).then_with(|| a.0.cmp(&b.0)));
+    let top = &sorted[..sorted.len().min(n)];
+    writeln!(w, "⏱ 최장 소요 top-{}:", top.len())?;
+    for (rank, (path, ms)) in top.iter().enumerate() {
+        writeln!(w, "  {}. {} — {}", rank + 1, path, fmt_ms(*ms))?;
+    }
+    Ok(())
+}
+
 /// Format the current wall-clock as RFC 3339 — used by `wire_ingest_progress`
 /// so every emitted event carries an `ts` field per §2.4a / the wire schema.
 pub(crate) fn now_rfc3339() -> anyhow::Result<String> {
@@ -348,4 +508,61 @@ mod tests {
         // well-formed RFC 3339 string.
         OffsetDateTime::parse(&s, &Rfc3339).expect("RFC 3339 round-trip");
     }
+
+    #[test]
+    fn abbreviate_path_passes_short_paths_through() {
+        assert_eq!(abbreviate_path("notes/foo.md"), "notes/foo.md");
+    }
+
+    #[test]
+    fn abbreviate_path_keeps_tail_with_ellipsis() {
+        let long = "a/very/deeply/nested/directory/structure/that/exceeds/the/budget/file.md";
+        let out = abbreviate_path(long);
+        assert!(out.starts_with('…'), "should be prefixed with ellipsis: {out}");
+        assert!(out.ends_with("file.md"), "should keep the filename tail: {out}");
+        // 48-char budget: 1 ellipsis + 47 tail chars.
+        assert_eq!(out.chars().count(), 48);
+    }
+
+    #[test]
+    fn write_slowest_summary_empty_writes_nothing() {
+        let mut buf = Vec::new();
+        write_slowest_summary(&mut buf, &[], 5).unwrap();
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn write_slowest_summary_sorts_desc_and_truncates() {
+        let timings = vec![
+            ("a.md".to_string(), 100),
+            ("b.png".to_string(), 5_000),
+            ("c.pdf".to_string(), 2_000),
+            ("d.md".to_string(), 50),
+        ];
+        let mut buf = Vec::new();
+        write_slowest_summary(&mut buf, &timings, 2).unwrap();
+        let out = String::from_utf8(buf).unwrap();
+        assert!(out.contains("top-2:"), "{out}");
+        // b (5s) ranks first, c (2s) second; a/d excluded.
+        let b_pos = out.find("b.png").expect("b.png present");
+        let c_pos = out.find("c.pdf").expect("c.pdf present");
+        assert!(b_pos < c_pos, "b before c: {out}");
+        assert!(!out.contains("a.md"), "a.md excluded by top-2: {out}");
+        assert!(out.contains("5.0s"), "b renders as 5.0s: {out}");
+    }
+
+    #[test]
+    fn write_slowest_summary_tie_breaks_by_path() {
+        let timings = vec![
+            ("z.md".to_string(), 1_000),
+            ("a.md".to_string(), 1_000),
+        ];
+        let mut buf = Vec::new();
+        write_slowest_summary(&mut buf, &timings, 5).unwrap();
+        let out = String::from_utf8(buf).unwrap();
+        assert!(
+            out.find("a.md").unwrap() < out.find("z.md").unwrap(),
+            "equal ms ties break alphabetically: {out}"
+        );
+    }
 }

crates/kebab-parse-image/src/ocr.rs

@@ -209,6 +209,13 @@ impl OllamaVisionOcr {
         self.max_pixels
     }
 
+    /// The Ollama model id this engine drives (e.g. `gemma4:e4b`).
+    /// Surfaced so the ingest progress display can name the model
+    /// running a slow OCR phase (`AssetPhase{phase:"ocr", model}`).
+    pub fn model(&self) -> &str {
+        &self.model
+    }
+
     fn build_prompt(&self, lang_hint: Option<&Lang>) -> String {
         let langs = if self.languages.is_empty() {
             "any".to_string()

crates/kebab-tui/src/ingest_progress.rs

@@ -160,7 +160,11 @@ fn apply_event(state: &mut IngestState, event: IngestEvent) {
         // per-asset counters, not sub-asset phase progress, so these are
         // no-ops here (the CLI / --json surfaces render them).
         | IngestEvent::AssetChunked { .. }
-        | IngestEvent::AssetTimings { .. } => {}
+        | IngestEvent::AssetTimings { .. }
+        // v0.26.0 slow-phase hint (ocr / caption / embed): the CLI bar uses
+        // it for a live phase message; the TUI status-bar reducer tracks only
+        // per-asset counters, so it's a no-op here.
+        | IngestEvent::AssetPhase { .. } => {}
     }
 }