fix(ui): keep customer copy action-oriented; add ADR 0237 accelerator boundary

ARCHITECTURE.md

@@ -25,6 +25,13 @@ recovery/equivalence checks; affected product paths fail closed during each
 cutover rather than substituting a local estimate. See
 `docs/doctoring/python-mathematical-compute-boundary-audit.md`.
 
+ADR 0237 also keeps accelerator deployment outside this repository. MLX runs
+as a native Apple-silicon inference service behind contextual-orchestrator;
+scientific CPU/CUDA/OpenCL profiles belong to TEPP or fast-mlsirm. RankWeave
+remains the dependency-free Python retrieval-fusion/evaluation owner behind its
+published contract. LineageWeave Compose therefore does not reserve devices or
+mount host drivers; its provider-neutral connectors consume versioned results
+and fail closed when an owning service is unavailable.
 ## Data flow
 
 ```mermaid

docs/adr/0237-accelerator-runtime-service-boundary.md

@@ -0,0 +1,100 @@
+# ADR 0237 — Accelerator runtimes stay behind owning service contracts
+
+**Decision status:** Accepted  
+**Date:** 2026-08-26  
+**Related:** ADR 0076, ADR 0083, ADR 0208
+
+## Context
+
+LineageWeave runs with Docker Compose on Linux, macOS, and Windows hosts, but
+does not own model inference or scientific computation. Adding MLX, CUDA, or
+OpenCL devices to its backend would duplicate upstream capability selection,
+couple the evidence API to host drivers, and make CPU-only installations less
+portable.
+
+The accelerator mechanisms are platform-specific. MLX targets Apple silicon's
+unified CPU/GPU memory. Docker Compose can reserve an NVIDIA GPU only when the
+host and daemon expose it, while the NVIDIA Container Toolkit injects host
+devices and driver mounts into a Linux container. OpenCL discovers vendor
+implementations through an installable-client-driver loader, so an image alone
+cannot prove that a compatible device and vendor driver are present.
+
+## Decision
+
+1. LineageWeave owns no MLX, CUDA, OpenCL, GPU, or scientific CPU runtime.
+   Its backend and frontend remain portable consumers of authenticated,
+   versioned service contracts.
+2. LLM, VISION, and embedding acceleration belongs to
+   contextual-orchestrator or a provider-neutral inference service registered
+   behind it. On Apple silicon, an MLX process runs natively as such a service;
+   LineageWeave does not pass Metal devices into its Linux VM or encode an MLX
+   URL, model, port, or chat template.
+3. TEPP and fast-mlsirm own their construct-specific scientific and
+   psychometric Rust cores. Their compute services may publish separate CPU and
+   accelerator deployment profiles: deterministic multithreaded CPU is the
+   portable required path; CUDA uses an explicit Compose GPU reservation plus
+   a compatible host driver/toolkit; OpenCL uses an explicitly mounted device
+   and matching vendor ICD. RankWeave remains the dependency-free Python owner
+   of retrieval fusion and evaluation behind its published contract; this ADR
+   neither changes its implementation language nor transfers psychometric
+   ownership to it. None of these profiles are added to LineageWeave Compose.
+4. LineageWeave connectors accept only the owner's provider-neutral envelope.
+   Persisted evidence records the owner, contract/model version, input/output
+   digest, execution-device class reported by the owner, convergence or
+   completion state, and uncertainty where the construct requires it. A device
+   label is provenance, not a quality score.
+5. Missing devices, drivers, ICDs, or owner services fail at the owning service
+   boundary. LineageWeave shows unavailable/failed status and the next valid
+   action; it never retries on a guessed backend, computes a Python substitute,
+   or claims GPU execution from configuration alone.
+
+```mermaid
+flowchart LR
+    LW[LineageWeave API and UI] -->|provider-neutral contract| CO[contextual-orchestrator]
+    LW -->|measurement contract| M[TEPP / fast-mlsirm service]
+    LW -->|retrieval-fusion contract| R[RankWeave]
+    CO --> N[Native MLX service on Apple silicon]
+    CO --> P[Remote or container inference provider]
+    M --> C[Deterministic multithreaded CPU]
+    M --> G[Owner CUDA or OpenCL profile]
+```
+
+## Considered alternatives
+
+- **Add accelerator profiles to LineageWeave Compose.** Rejected because this
+  repository does not own the computation and cannot validate host drivers for
+  another service's algorithm.
+- **Run every accelerator natively.** Rejected because CUDA containers are a
+  supported owner deployment when their host prerequisites are explicit.
+- **Use CPU fallback inside LineageWeave.** Rejected because it would reproduce
+  the formula on the wrong side of the contract.
+
+## Consequences and acceptance
+
+- LineageWeave Compose remains CPU-portable and contains no device reservation.
+- TEPP and fast-mlsirm bear deployment and recovery-test work for every
+  advertised scientific-compute profile. RankWeave retains its own retrieval
+  fusion/evaluation conformance contract.
+- A scientific-compute integration is accepted only when its owning repository
+  proves the same versioned synthetic input on deterministic CPU and each
+  advertised accelerator, reports bounded numerical tolerance and device
+  provenance, and LineageWeave proves malformed, mismatched, and unavailable
+  envelopes fail closed without exposing implementation details in customer
+  copy.
+- Native MLX availability is verified at contextual-orchestrator's provider
+  boundary; CUDA/OpenCL availability is verified in the compute owner's health
+  and conformance evidence. A Compose declaration by itself is insufficient.
+
+## References (APA 7th)
+
+Docker, Inc. (2026). *Run Docker Compose services with GPU access*.
+https://docs.docker.com/compose/how-tos/gpu-support/
+
+Khronos Group. (2026). *OpenCL registry*.
+https://registry.khronos.org/OpenCL/
+
+MLX Contributors. (2026). *Unified memory*. MLX documentation.
+https://ml-explore.github.io/mlx/build/html/usage/unified_memory.html
+
+NVIDIA Corporation. (2026). *NVIDIA Container Toolkit architecture overview*.
+https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/arch-overview.html

docs/product-technical-gap-baseline.md

@@ -384,6 +384,7 @@ this file per §3.5 of the prior snapshot).
 | Frontend delivery performance | #644 implements a native dynamic-import boundary for conditional workspace surfaces and retains accessible loading/error states; exact-head checks passed but the PR is not protected-main evidence | Merge #644 normally, rebuild the protected-main production bundle, and retain the measured chunk inventory rather than raising the warning limit |
 | External integrations | Search, Zotero, calendar, Keyverse, orchestrator, RankWeave, ThreadWeave, TEPP, DiskSage, wardnet | Provider conformance, failure/reconciliation behavior, and provenance-bearing integration evidence |
 | MSA / modular reuse | LineageWeave must run standalone and as a consumer of org packages | Do not reimplement RankWeave/TEPP/orchestrator/ThreadWeave/Keyverse; fix upstream and PR there |
+| Accelerator runtime ownership | ADR 0076/0208 already prohibit local model and mathematical ownership; ADR 0237 now defines MLX as a native orchestrator-side service and TEPP/fast-mlsirm CUDA/OpenCL/CPU profiles as scientific-compute-owner deployments, so LineageWeave Compose remains device-neutral. RankWeave remains the dependency-free Python retrieval-fusion/evaluation owner behind its published contract | TEPP and fast-mlsirm must publish deterministic CPU recovery plus conformance evidence for every advertised CUDA/OpenCL profile; contextual-orchestrator must prove native MLX availability through its provider-neutral health/contract boundary. LineageWeave accepts only versioned, provenance-bearing envelopes and fails closed when the owner is unavailable |
 | Product contract authority | The current LineageWeave PRD records exact-case ecosystem authorities. TEPP, fast-mlsirm, keyverse, ThreadWeave, and RankWeave PR #41 have standalone PRDs; RankWeave's remains unmerged. contextual-orchestrator, disksage, and wardnet still rely on product/architecture documents, and naruon has only a scoped Topic Intelligence PRD | Keep ADRs normative, preserve canonical repository case in machine references, land the pending PRDs, and add standalone PRDs in each remaining owning repository before cross-product release claims exceed its documented boundary |
 | Release quality | PR #660 is now on protected `main`; its pre-merge full Python suite passed 1,352 tests with 17 skips, but release-wide frontend, Storybook, security, browser, and runtime acceptance remain unproven on one exact protected head | Repository-wide coverage, docstrings, Storybook, security, browser, and release evidence on one exact head |
 | PII | Masking would paralyze the product; ADR 0001 forbids identifying artifacts in git | ABAC + authorized runtime; synthetic fixtures in git; no mask-in-place that drops names the operator must read |

frontend/src/App.test.tsx

@@ -2183,7 +2183,7 @@ describe("App, authenticated", () => {
       "hint-disclosure",
     );
 
-    const authorSection = screen.getByRole("region", { name: "Source author evidence" });
+    const authorSection = screen.getByRole("region", { name: "Author context" });
     expect(within(authorSection).getByText("AUTH-HINT · Hint only").closest("details")).toHaveClass(
       "hint-disclosure",
     );
@@ -2456,13 +2456,13 @@ describe("App, authenticated", () => {
     await userEvent.click(await screen.findByRole("button", { name: "View post: Public post" }));
 
     await waitFor(() => expect(screen.getByText("이것은 요약입니다.")).toBeInTheDocument());
-    const provenance = screen.getByText("Evidence provenance").closest("details");
+    const provenance = screen.getByText("Why this item is listed").closest("details");
     expect(provenance).not.toBeNull();
     expect(provenance).not.toHaveAttribute("open");
-    await userEvent.click(screen.getByText("Evidence provenance"));
-    expect(screen.getByText(/Ontology class:/)).toBeInTheDocument();
-    expect(screen.getByText(/Extraction source: Semantic extraction/)).toBeInTheDocument();
-    expect(screen.getByText(/Evidence field: Stored semantic evidence/)).toBeInTheDocument();
+    await userEvent.click(screen.getByText("Why this item is listed"));
+    expect(screen.getByText(/Category:/)).toBeInTheDocument();
+    expect(screen.getByText(/How this item was found: Semantic extraction/)).toBeInTheDocument();
+    expect(screen.getByText(/Recorded evidence: Stored semantic evidence/)).toBeInTheDocument();
     expect(screen.queryByText("contextual_orchestrator_semantic")).not.toBeInTheDocument();
     expect(screen.queryByText("https://contextualwisdomlab.github.io/LineageWeave/ontology#Project")).not.toBeInTheDocument();
     expect(screen.getByText("첫 번째 이벤트")).toBeInTheDocument();
@@ -2696,7 +2696,11 @@ describe("App, authenticated", () => {
     await userEvent.click(await screen.findByRole("button", { name: /verify against web search/i }));
 
     await waitFor(() =>
-      expect(screen.getByText("Verification unavailable (search is not configured).")).toBeInTheDocument(),
+      expect(
+        screen.getByText(
+          "Verification is unavailable because public search is not configured yet. Ask an administrator to enable it, then retry.",
+        ),
+      ).toBeInTheDocument(),
     );
     expect(screen.queryByText(/HTTP 503/)).not.toBeInTheDocument();
     expect(screen.queryByRole("button", { name: /verify against web search/i })).not.toBeInTheDocument();
@@ -3088,11 +3092,15 @@ describe("App, authenticated", () => {
     );
   });
 
-  it("names RankWeave unavailability on home rankings instead of inventing a fused score", async () => {
+  it("names rankings unavailability on home rankings instead of inventing a score", async () => {
     stubBackend();
     render(<App />);
 
-    expect(await screen.findByText("Rankings · RankWeave not available")).toBeInTheDocument();
+    expect(
+      await screen.findByText(
+        "Rankings are not available right now. Reopen this post later to load them.",
+      ),
+    ).toBeInTheDocument();
     expect(screen.queryByText("Pricing renegotiation: revised quote sent")).not.toBeInTheDocument();
   });
 
@@ -3196,11 +3204,11 @@ describe("App, authenticated", () => {
       name: /open ranking: public post/i,
     });
     expect(rankingButton).toHaveTextContent("Public post");
-    expect(rankingButton).toHaveTextContent("Rankings · rankweave");
+    expect(rankingButton).toHaveTextContent("Rankings");
     expect(rankingButton).toHaveTextContent("rank 1");
     expect(
       screen.getByText(
-        "RankWeave fused newest-first and title-overlap ranks. This is not a calibrated score.",
+        "Rankings combine newest-first and title-overlap evidence and are not calibrated scores. Open a ranked post to see its evidence.",
       ),
     ).toBeInTheDocument();
     expect(
@@ -3249,7 +3257,7 @@ describe("App, authenticated", () => {
     expect(list).toHaveTextContent("TEPP measurement · Failed · Demo Corp");
     expect(list).toHaveTextContent("Period report · Succeeded · Demo Corp");
     expect(list).toHaveTextContent(
-      "Open this run to see why it failed, then connect the measurement service and re-run.",
+      "Open this run to see why it failed, then retry with the latest available records.",
     );
     expect(list).toHaveTextContent("3 documents");
     expect(list).not.toHaveTextContent("postgresql://");
@@ -3432,8 +3440,9 @@ describe("App, authenticated", () => {
     );
     expect(lineageButton).not.toHaveTextContent("measurement service");
     expect(teppButton).toHaveTextContent(
-      "Open this run to see why it failed, then connect the measurement service and re-run.",
+      "Open this run to see why it failed, then retry with the latest available records.",
     );
+    expect(teppButton).not.toHaveTextContent("measurement service");
     expect(teppButton).not.toHaveTextContent("reconstruction");
   });
 
@@ -3721,7 +3730,7 @@ describe("App, authenticated", () => {
     ).not.toBeInTheDocument();
     expect(
       await screen.findByText(
-        "No posts were available at this cutoff for the period report. Open a later run, or ask an administrator to capture a newer snapshot.",
+        "No posts were available at this cutoff for the period report. Open a later run or retry after a newer snapshot is available.",
       ),
     ).toBeInTheDocument();
   });