ltcai 10.6.0 → 10.6.2

package/README.md

@@ -11,7 +11,7 @@
 [![CI Status](https://github.com/TaeSooPark-PTS/LatticeAI/actions/workflows/ci.yml/badge.svg?branch=main)](https://github.com/TaeSooPark-PTS/LatticeAI/actions/workflows/ci.yml)
 [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
 
-![v10.6.0 Living Brain walkthrough](output/release/v10.6.0/gifs/v10.6.0-living-brain-walkthrough.gif)
+![v10.6.2 Living Brain walkthrough](output/release/v10.6.2/gifs/v10.6.2-living-brain-walkthrough.gif)
 
 Chat, files, folders, notes, and web pages all flow into one durable knowledge
 graph on your computer. Any model — local MLX or cloud — can speak with that
@@ -24,10 +24,10 @@ memory. Nothing leaves your machine without explicit consent.
 
 | | |
 | --- | --- |
-| **Chat with a Brain that remembers** — every conversation grows durable, source-linked memory ![Brain Chat](output/release/v10.6.0/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v10.6.0/screenshots/05-memory-graph.png) |
-| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v10.6.0/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v10.6.0/screenshots/12-review-center.png) |
-| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v10.6.0/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v10.6.0/screenshots/10-admin-console.png) |
-| **Watch a file become memory** — three named steps, not a pipeline diagram ![Material to memory](output/release/v10.6.0/screenshots/11-knowledge-journey.png) | **Say how much it may do alone** — one dial in plain words; dangerous actions stay blocked either way ![Settings](output/release/v10.6.0/screenshots/08-system.png) |
+| **Chat with a Brain that remembers** — every conversation grows durable, source-linked memory ![Brain Chat](output/release/v10.6.2/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v10.6.2/screenshots/05-memory-graph.png) |
+| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v10.6.2/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v10.6.2/screenshots/12-review-center.png) |
+| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v10.6.2/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v10.6.2/screenshots/10-admin-console.png) |
+| **Watch a file become memory** — three named steps, not a pipeline diagram ![Material to memory](output/release/v10.6.2/screenshots/11-knowledge-journey.png) | **Say how much it may do alone** — one dial in plain words; dangerous actions stay blocked either way ![Settings](output/release/v10.6.2/screenshots/08-system.png) |
 
 ## Why Lattice AI
 
@@ -58,62 +58,69 @@ First-run flow — wake the Brain, pick the owner, load a recommended model:
 
 | | | |
 | --- | --- | --- |
-| ![Login](output/release/v10.6.0/screenshots/01-login.png) | ![Model install](output/release/v10.6.0/screenshots/03-install-load-progress.png) | ![Model library](output/release/v10.6.0/screenshots/07-model-library.png) |
+| ![Login](output/release/v10.6.2/screenshots/01-login.png) | ![Model install](output/release/v10.6.2/screenshots/03-install-load-progress.png) | ![Model library](output/release/v10.6.2/screenshots/07-model-library.png) |
 
 Screenshot index and capture notes:
-[output/release/v10.6.0/SCREENSHOT_INDEX.md](output/release/v10.6.0/SCREENSHOT_INDEX.md)
+[output/release/v10.6.2/SCREENSHOT_INDEX.md](output/release/v10.6.2/SCREENSHOT_INDEX.md)
 
 ## Current Release
 
-The current release is **10.6.0 — Promoted Panels**:
-
-10.5.0 changed the words on each screen. 10.6.0 changes where things sit. Every
-main screen used to open as a row of equal tabs, which asks a first-time user to
-choose before they know what the choices are. Each screen now opens on the panel
-that answers the question that brought you there, and everything else moves
-below it. No feature was removed — every panel is still on the page, it just no
-longer competes for the top of it.
-
-- **Each screen leads with one panel instead of a row of equals.** Capture opens
-  on a single 자료 추가하기 card holding all three ways to add material —
-  파일 올리기 / 폴더 연결하기 / 웹페이지 저장하기 — as a choice inside one card
-  rather than three tabs, with progress and connected folders moved to a quieter
-  second row. Work opens on 검토함, what is waiting on you, instead of on the
-  empty goal composer. The model library answers "which model is running, and
-  can I switch it" in a card above the tabs.
-- **The Brain home is one card, not five stacked blocks.** Greeting, composer,
-  the add-material row and the autonomy dial now sit inside a single bordered
-  station that lifts when you type anywhere in it. The Brain artwork shrank so
-  that it introduces the composer rather than headlining the screen.
-- **Everyday and management destinations stopped being one list.** 대화 · 자료 ·
-  기억 stay in the primary nav; 작업 · AI 모델 · 설정 became topbar links at
-  desktop widths and fold into the menu below that — built from one array so the
-  two copies cannot drift, and shown by one breakpoint so they can never both
-  appear or both vanish.
-- **The knowledge graph became a subview rather than a third tab.** Memory opens
-  on search, with 연결 지도 열기 one button away and a labelled way back.
-- **Seven settings tabs became three named groups** — 나와 작업공간 · 내 데이터
-  보관 · 동작 방식과 연결 — so the row reads as a short list of decisions.
-- **Links that named a screen now open that screen.** `#/act/review` and its
-  siblings landed on the Brain home for every caller that emitted them — the
-  command palette, the daily briefing — because nothing parsed the
-  `<screen>/<tab>` form they had always used. The palette also read a private
-  second copy of the destination list, so 작업 opened a different screen
-  depending on whether you clicked it or searched for it. One list, one parser.
-- **Guarded, not asserted.** The visual sweep walks ten viewport widths and
-  fails if a management link is visible in both places at once or in neither, if
-  a navigation landmark is unnamed or shares its name with another, or if the
-  topbar overflows.
+The current release is **10.6.2 — Ask First**:
+
+10.6.1 rebuilt five screens, and the Brain home is the one it only half
+rebuilt: the order changed, the shape did not. A large Brain and a centred
+headline still ran down the middle of the screen before the box you type into,
+and the three suggestions still sat between that box and its own toolbar — so
+the alternative to typing wore the same border as typing, and the composer was
+separated from its own controls by an unrelated block. This release splits the
+screen into two surfaces. Nothing was removed; the second choice stopped
+sharing a card with the first.
+
+- **The greeting introduces the composer instead of headlining the screen.**
+  The Brain and the greeting were a centred column above the input; they are a
+  compact banner across the top of the card now — a small Brain beside the title
+  and the memory count, on a tinted strip closed by a hairline. The Brain is
+  still the button that opens the memory map, and the banner stacks back into a
+  centred column on a phone.
+- **One card is the first move, and only the first move.** The station reads
+  greeting → the box you type into → one row for 자료 추가 and 혼자 해도 되는 일,
+  with nothing between the box and its own controls.
+- **What to try is a card of its own.** The three suggestions moved out of the
+  station onto a second surface below it, where they read as an offer rather
+  than a step. That card is a named `<section>`, so a screen reader announces it
+  as a region — the label used to sit on a plain `<div>`, where the browser
+  discards it.
+- **The empty state is designed rather than inherited.** A Brain with nothing to
+  suggest yet — every Brain on its first day — shows starter pills instead of
+  suggestion cards. That row had never been styled for this screen: it borrowed
+  a pill height drawn for live conversations and sat noticeably taller than the
+  cards it stands in for. Both branches read as peers now, and a pill fills the
+  composer rather than sending a question you did not choose to ask.
+- **Three bugs that only moving things could find.** The Brain's halo is written
+  as an inline `box-shadow`, so no stylesheet could shrink it — a 60px glow
+  stayed wrapped around a 58px Brain until the blur became a CSS variable the
+  host can set (every other screen keeps its old value through the fallback).
+  Clipping the card would have hidden the 노트 / 웹 popover and turned the card
+  into a scroll container that scrolls the greeting away on focus, so the banner
+  clips its own corners instead. And the reduced-motion rule that cancels the
+  card lift had to follow the cards to their new selector, or a reader who asked
+  the system to hold still would have got the animation back.
+- **Guarded, not asserted.** Unit tests hold the split itself — station, deck
+  and footer as three siblings, with the suggestions provably *not* inside the
+  station. Browser tests hold what only a browser can see: the popover opens
+  unclipped, the cards fill their card, they collapse to chips under 900px and
+  stay visible under 760px, reduced motion is honoured, the empty-state pill row
+  renders, and the halo is measured against the Brain it surrounds.
 
 Release notes: [RELEASE.md](RELEASE.md) · Full history: [docs/CHANGELOG.md](docs/CHANGELOG.md)
 
-Expected artifacts for 10.6.0 release must use exact filenames:
+Expected artifacts for 10.6.2 release must use exact filenames:
 
-- `dist/ltcai-10.6.0-py3-none-any.whl`
-- `dist/ltcai-10.6.0.tar.gz`
-- `ltcai-10.6.0.tgz`
-- `dist/ltcai-10.6.0.vsix`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_10.6.0_aarch64.dmg`
+- `dist/ltcai-10.6.2-py3-none-any.whl`
+- `dist/ltcai-10.6.2.tar.gz`
+- `ltcai-10.6.2.tgz`
+- `dist/ltcai-10.6.2.vsix`
+- `src-tauri/target/release/bundle/dmg/Lattice AI_10.6.2_aarch64.dmg`
 
 Do not use wildcard artifact uploads. Package registry publishing remains owner-run.
 
@@ -148,6 +155,8 @@ See [ARCHITECTURE.md](ARCHITECTURE.md) for details and
 
 | Version | Theme |
 | --- | --- |
+| 10.6.2 | Ask First |
+| 10.6.1 | First Things |
 | 10.6.0 | Promoted Panels |
 | 10.5.0 | Everyday Words |
 | 10.4.0 | Named Ground |

package/docs/CHANGELOG.md

@@ -4,6 +4,132 @@ The top entry is either the current unreleased main-branch work or the current
 release line. Older entries are historical and may describe behavior as it
 existed at that release.
 
+## [10.6.2] - 2026-08-03
+
+The Brain home, re-shaped rather than re-ordered. 10.6.1 gave this screen a new
+order and left its shape alone: a 5.4rem organism and a centred headline ran
+down the middle before the composer, and the suggestion strip sat between the
+composer and its own toolbar. It is two surfaces now — a station holding the
+first move, and a deck below it holding the alternatives. No feature was
+removed. Backend behaviour is unchanged; `frontend/openapi.json` differs from
+10.6.1 only in `info.version`.
+
+### Changed
+- **The greeting is a banner, not a headline.** `.brain-home-station > .brain-hero`
+  is a `row` instead of a `column`: organism 5.4rem → 3.2rem beside the title and
+  the memory counts (`.brain-hero-header-text`), on a tinted strip with a
+  hairline foot. It stacks back to a centred column under 640px, and the
+  short-window rule now shrinks the organism to 2.6rem rather than growing it.
+- **The station holds the first move and nothing else.** Greeting → composer →
+  one toolbar (`자료 추가` left, `혼자 해도 되는 일` right) with nothing between
+  the composer and its own controls.
+- **The suggestions are their own card.** `.brain-secondary-deck` is a sibling
+  of the station under `.brain-centered-home`, which now carries three direct
+  children: station, deck, quiet row. The stage widened 44rem → 50rem.
+- **The empty branch is styled for this screen.** With no `suggested_questions`,
+  the deck renders `.brain-prompt-pills-row`, scoped and sized here (notably
+  `min-height: 0`, against `conversation.css`'s 2.65rem pill floor) so both
+  branches read as peers. A pill fills the composer instead of sending.
+
+### Fixed
+- **The Brain's halo could not be scaled by any stylesheet.** `LivingBrain`
+  writes the aura's `box-shadow` inline, so a `.brain-hero-organism .brain-aura`
+  rule had no effect and a 60px blur stayed around a 58px organism, clipped into
+  a flat smudge. The blur reads `var(--aura-blur, 60px)`; the banner sets 14px.
+  Every other host is pixel-identical through the fallback.
+- **`aria-label` on a plain `<div>` was being discarded.** The suggestion block
+  reached assistive tech unnamed. The label moved to the deck's `<section>`
+  (named ⇒ `region`); the inner strip carries none.
+- **Reduced motion stopped covering the suggestion cards.** `affordance.css`'s
+  hover-lift and `prefers-reduced-motion` selectors still named
+  `.brain-home-prompt-strip > button`, which matches nothing after the grid; they
+  are re-pointed at `.brain-prompt-grid button`. The `white-space` opt-in entry
+  is deliberately not re-added — `home-simple.css` flips those cards to `nowrap`
+  chips under 900px and a matching selector in the last-loaded sheet would win
+  that media rule.
+- **The station must not clip.** `overflow: hidden` there would hide the capture
+  popover (anchored to the toolbar, opening below it) and make the card a scroll
+  container that scrolls the greeting away on focus. The station stays visible;
+  the banner uses `overflow: clip` and rounds its own top corners.
+
+### Tests
+- `BrainHome.test.tsx`: station/deck/quiet are direct siblings of the stage in
+  order; the station contains neither the strip nor the deck; the deck is a named
+  `region` and the inner strip is unlabelled; the starter-pill branch renders.
+- `tests/visual/v3.spec.js`: the popover opens unclipped without scrolling the
+  station; the grid fills ≥98% of the deck's inner width; the cards stay visible
+  and become chips at 900/760/640/420px; reduced motion drops both `transform`
+  and `transition` while hover still answers in colour; the aura's blur is
+  measured against the organism and its box against the banner.
+
+## [10.6.1] - 2026-08-03
+
+The second half of the layout rebuild. 10.6.0 promoted one panel per main screen
+and left five screens untouched: sign-in, the recommended model, the Brain home,
+the automation runs list, and the review center. Each of those now opens on the
+one thing it exists for, with everything else grouped beneath it. No feature was
+removed. Backend behaviour is unchanged; `frontend/openapi.json` differs from
+10.6.0 only in `info.version`.
+
+### Changed
+- **Login is one card.** The three-card promise bar moved from between the
+  greeting and the form to a quiet hairline strip below it
+  (`.ritual-promise.is-quiet`), leaving the form as the only raised surface. The
+  password-stays-local note and the existing-Brain note now read after the
+  submit button instead of between the last field and it.
+- **The recommendation screen names its top pick once.** The duplicate CTA above
+  the list is gone; rank, size, name, reason, time estimate and the action live
+  in one `.ritual-primary-hero-card`, with the other two models under a labelled
+  `다른 선택지` grid as `.ritual-model-card.is-compact`. The footer splits 뒤로 to
+  one end and the hint + 모델 없이 Brain 열기 to the other.
+- **The Brain home leads with the composer.** It gained its own bordered wrapper
+  and focus ring; the suggestion strip moved directly under it and became a card
+  grid that shows each suggestion's detail line (previously `title`-only); the
+  add-material dock and the autonomy dial dropped to the station floor as one
+  toolbar; the past-conversation count reads as a badge.
+- **The runs tab stacks by urgency.** `승인함` moved from the bottom to the top
+  with an attention treatment (`.data-panel.is-attention`), installed automations
+  became visible from this tab with their last run — mode, result, timestamp,
+  summary — inside the card, and the agent and workflow run tables moved to the
+  end as history.
+- **A review item is evidence beside a decision.** `ReviewCard` splits into a
+  7/5 grid: proposal diff, snooze state, risk/class/tool provenance and the
+  technical `<details>` on the left; an always-visible decision panel with
+  approve / reject / snooze / run-now and the reject reason on the right. Items
+  with no evidence render as a single column instead of an empty left half.
+
+### Accessibility
+- Login: `<label for>` on every field, form named by the page heading,
+  `aria-invalid` + `aria-describedby` wired to the error only once it exists.
+- Recommendation: hero card labelled by the model name it recommends.
+- Brain home: greeting is the page `<h1>`; station header and footer scoped so
+  they do not register as page landmarks; suggestion strip and station toolbar
+  carry accessible names.
+- Review: each item is an `<article>` named by its `<h3>`, the decision block is
+  an `<h4>`, the approve/reject cluster is a named `role="group"`, and the status
+  and source filters have visible captions referenced by `aria-labelledby`.
+
+### Fixed
+- `p-6` on the Brain home stage added padding on top of children that already
+  pad themselves and pushed the quiet shelves under the fixed mobile nav, where
+  the tap hit the nav. The project's stylesheets are unlayered and therefore beat
+  Tailwind's `@layer utilities` for properties they set — but not for properties
+  they leave unset, which is why this one applied.
+  `frontend/src/styles/cssLayering.test.ts` asserts the project ships no
+  `@layer`, which is what makes that rule hold.
+- The installed-automations empty state pointed at "the suggestions above" from
+  a tab that has none above it; `act.installed.empty` now names the 레시피 tab in
+  both languages.
+
+### Tests
+- New: `LoginScreen.test.tsx`, `RecommendationScreen.test.tsx`,
+  `BrainHome.test.tsx`, `ReviewCard.test.tsx`, `ReviewInbox.test.tsx`,
+  `cssLayering.test.ts`; `Act.test.tsx` gained runs-tab ordering coverage.
+- `tests/visual/mock_server.cjs` serves `/api/proposals/counts`, two installed
+  automations with last-run detail, and a `change_proposal` review item with a
+  real diff — without them the promoted approval block and the new evidence
+  column are captured empty.
+
 ## [10.6.0] - 2026-08-03
 
 A layout rebuild, not a copy pass. Every main screen used to open as a row of

package/docs/COMMUNITY_AND_PLUGINS.md

@@ -1,6 +1,6 @@
 # Community And Plugins
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 LatticeAI defines the path from a strong local-first framework (8.4.0
 action-aware baseline, 8.5.0 registry+DI hardening, 8.6.0 capture/navigation

package/docs/DEVELOPMENT.md

@@ -3,7 +3,7 @@
 > **Status: canonical** — current contributor guidance, kept in sync with the
 > current release.
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 This document is for contributors working on the local-first Digital Brain
 codebase. Product positioning and quick start stay in `README.md`; release

package/docs/ONBOARDING.md

@@ -1,6 +1,6 @@
 # Lattice AI Onboarding
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 The first-run goal is a five-minute path from "I opened the app" to "my Brain
 has a source, a question, and proof." This page is the product contract behind

package/docs/OPERATIONS.md

@@ -1,4 +1,4 @@
-# Lattice AI — Operations Guide (v10.6.0)
+# Lattice AI — Operations Guide (v10.6.2)
 
 > **Status: canonical** — kept in sync with the current release. Storage layout
 > below reflects the SQLite live Brain store and workspace scoping, not the

package/docs/TRUST_MODEL.md

@@ -1,6 +1,6 @@
 # Lattice AI Trust Model
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 Lattice AI is local-first, explicit about external communication, and honest
 when a capability is unavailable.

package/docs/WHY_LATTICE.md

@@ -1,6 +1,6 @@
 # Why Lattice AI Exists
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 **Lattice AI is a local-first Digital Brain that keeps your knowledge durable
 across any AI model.**

package/docs/kg-schema.md

@@ -1,6 +1,6 @@
 # Knowledge Graph Schema
 
-Current release: **10.6.0 — Promoted Panels**.
+Current release: **10.6.2 — Ask First**.
 
 명세 출처: `lattice_ai_full_spec.pptx` 슬라이드 20·21·22
 구현: `lattice_brain/graph/schema.py`

package/lattice_brain/__init__.py

@@ -26,7 +26,7 @@ from .storage import (
     storage_from_env,
 )
 
-__version__ = "10.6.0"
+__version__ = "10.6.2"
 
 __all__ = [
     "AgentRuntime",

package/lattice_brain/graph/_kg_contract.py

@@ -158,6 +158,21 @@ class KnowledgeGraphCore:
     ) -> List[str]:
         raise NotImplementedError
 
+    # ── retrieval_vector.py: derived vector index ────────────────────────────
+    def rebuild_vector_index(
+        self,
+        *,
+        full: bool = False,
+        include_nodes: bool = True,
+        include_chunks: bool = True,
+    ) -> Dict[str, Any]:
+        """Re-derive the vector index; also records the embedder fingerprint."""
+        raise NotImplementedError
+
+    def vector_freshness(self) -> Dict[str, Any]:
+        """``{status, pending_items, total_items, detail}`` — never raises."""
+        raise NotImplementedError
+
     # ── retrieval_reads.py: workspace scoping ────────────────────────────────
     def filter_scoped_nodes(
         self,

package/lattice_brain/graph/provenance.py

@@ -217,8 +217,10 @@ class KnowledgeGraphProvenanceMixin(_Core):
         include_legacy_global: bool = False,
     ) -> Dict[str, Any]:
         """Raw, lossless logical export of the graph (nodes/edges/chunks/sources/
-        provenance). Vector embeddings are intentionally omitted — they are
-        re-derived on import — so the artifact stays portable and small. Use
+        provenance). Vector embeddings are intentionally omitted — the importer
+        re-derives them with *its own* embedder (:meth:`import_graph_data`
+        embeds on write and then reindexes, reporting the result under
+        ``index``) — so the artifact stays portable and small. Use
         :meth:`backup_database` for a faithful binary copy incl. embeddings.
 
         ``workspace_id`` REALLY filters (v4): the artifact contains only nodes
@@ -274,15 +276,68 @@ class KnowledgeGraphProvenanceMixin(_Core):
         data["counts"] = {k: len(v) for k, v in data.items()}
         return data
 
+    def _reindex_after_import(self) -> Dict[str, Any]:
+        """Bring the vector index in line with what was just imported.
+
+        A logical artifact carries no embeddings (see :meth:`export_graph_data`)
+        so they have to be re-derived on this machine. The write door
+        (``_upsert_node`` / ``_upsert_chunk``) already embeds inline, which
+        makes this pass a *verification* in the common case — but "the index is
+        consistent after an import" has to be a guarantee the import makes, not
+        an accident of where the embedding call happens to live. It also
+        records the embedder fingerprint, which the inline write path never
+        does; without it ``index_status()`` reports ``recorded: None`` and the
+        ``stale_embedder`` honesty signal is dead after every import.
+
+        Never raises. The graph rows are already committed at this point, so an
+        embedding provider that dies here means "content imported, recall
+        degraded" — reported with the repo's existing ``vector_freshness``
+        vocabulary (``ready`` / ``pending`` / ``stale_embedder`` /
+        ``unavailable``) plus an explicit ``degraded`` flag — never a rollback
+        of work that already landed, and never a silent success.
+        """
+        reindexed = 0
+        try:
+            outcome = self.rebuild_vector_index(full=False) or {}
+            reindexed = int(outcome.get("items_indexed") or 0)
+        except Exception as exc:  # noqa: BLE001 — the import already committed
+            return {
+                "status": "unavailable",
+                "degraded": True,
+                "detail": (
+                    "imported content is stored, but the vector index could not "
+                    f"be rebuilt ({exc}); search stays lexical-only for the new "
+                    "items until a rebuild succeeds"
+                ),
+                "reindexed_items": 0,
+            }
+        report = self.vector_freshness() or {}
+        status = str(report.get("status") or "unavailable")
+        return {
+            "status": status,
+            "degraded": status != "ready",
+            "detail": report.get("detail"),
+            "pending_items": int(report.get("pending_items") or 0),
+            "total_items": int(report.get("total_items") or 0),
+            "reindexed_items": reindexed,
+        }
+
     def import_graph_data(
         self, data: Dict[str, Any], *, mode: str = "merge", dry_run: bool = False
     ) -> Dict[str, Any]:
         """Import a logical export back into the store.
 
         ``mode='merge'`` upserts on top of existing data (id collisions update);
-        ``mode='replace'`` clears the graph first. ``dry_run=True`` reports the
-        plan without writing. Refuses artifacts from a NEWER graph schema than
-        this build.
+        ``mode='replace'`` clears the graph first (including the derived vector
+        rows — the re-import re-embeds them inside the same transaction, so a
+        failed artifact rolls back to the previous index rather than an empty
+        one). ``dry_run=True`` reports the plan without writing. Refuses
+        artifacts from a NEWER graph schema than this build.
+
+        A completed import carries an ``index`` block reporting the state of
+        the re-derived vector index (see :meth:`_reindex_after_import`); when
+        ``index["degraded"]`` is true, retrieval for the imported scope is
+        lexical-only until the index is rebuilt.
         """
         nodes = data.get("nodes") or []
         edges = data.get("edges") or []
@@ -418,6 +473,7 @@ class KnowledgeGraphProvenanceMixin(_Core):
                     ),
                 )
         plan["imported"] = True
+        plan["index"] = self._reindex_after_import()
         return plan
 
     def backup_database(self, dest_path) -> Path:

package/lattice_brain/graph/retrieval.py

@@ -454,17 +454,25 @@ class KnowledgeGraphRetrievalMixin(_Core):
         mode = "hybrid"
         detail: Optional[str] = None
         vector_matches: List[Dict[str, Any]] = []
+        vector_recall: Optional[Dict[str, Any]] = None
         vector_fn = getattr(self, "vector_search", None)
         if not callable(vector_fn):
             mode = "lexical_only"
             detail = "vector search is not available on this store"
         else:
             try:
-                vector_matches = list(
-                    (vector_fn(search_query, limit=vec_fetch, min_score=min_vector_score) or {}).get(
-                        "matches", []
-                    )
+                vector_payload = (
+                    vector_fn(search_query, limit=vec_fetch, min_score=min_vector_score)
+                    or {}
                 )
+                vector_matches = list(vector_payload.get("matches", []))
+                # Partial recall must reach the caller: the vector channel can
+                # only score a capped slice of a large index (see
+                # retrieval_vector.vector_search), and a fused answer built on
+                # a truncated scan is not the same claim as a complete one.
+                recall = vector_payload.get("recall")
+                if isinstance(recall, dict) and recall.get("truncated"):
+                    vector_recall = dict(recall)
             except Exception as exc:  # noqa: BLE001 — degrade, never fail the search
                 mode = "lexical_only"
                 detail = f"vector index unavailable: {exc}"
@@ -619,6 +627,10 @@ class KnowledgeGraphRetrievalMixin(_Core):
         }
         if vector_degraded is not None:
             result["vector_degraded"] = vector_degraded
+        if vector_recall is not None:
+            result["vector_recall"] = vector_recall
+            if vector_degraded is None:
+                result["vector_degraded"] = "partial_recall"
         return result
 
     def context_for_query(