fix: follow-up batch (learning, small-GPU loading, chat history, CI cost, e2e layout)

.github/WORKFLOW_README.md

@@ -32,11 +32,12 @@ This repository is configured with automated dependency updates and release buil
 
 ### 3. CI (`.github/workflows/ci.yml`)
 
-- **Job (`ci`)**: on pull requests, `npm ci` then **`npm run ci`** on Ubuntu, Windows, and macOS
-  (format, lint, typecheck, production audit, commercial compliance verify, `test:coverage`,
-  `generate:channels:check`, the full `verify:*` suite, and `build` — see `package.json` for the
-  exact step list).
-- **Job (`e2e-smoke`)**: Playwright smoke specs on Ubuntu and Windows.
+- **Job (`ci`)**: `npm ci` then **`npm run ci`** (format, lint, typecheck, production audit,
+  commercial compliance verify, `test:coverage`, `generate:channels:check`, the full `verify:*`
+  suite, and `build` — see `package.json` for the exact step list). Pull requests run it on Ubuntu
+  and Windows; pushes add macOS.
+- **Job (`e2e-smoke`)**: Playwright smoke specs on Ubuntu and Windows, on pushes only. macOS and the
+  smoke are push-only to keep Actions minutes down, so run the smoke specs locally before pushing.
 - **Stricter local gate** (includes Electron-linked native check): `npm run local:gate` (=
   `verify:installed-electron` + `npm run ci` + `verify:native-modules`).
 - Triggered on pull requests and pushes to main/master/develop/RegressionFixes

.github/workflows/ci.yml

@@ -1,12 +1,16 @@
 # Pull requests: run the same gate as packaged in `npm run ci` (format, lint, audit,
-# compliance verify, tests + coverage, IPC contracts, production webpack build).
-# PR CI also verifies Electron-linked native modules load on all three OSes. The
-# macos-latest runner builds x64 while production ships arm64 — the x64 run still
-# exercises the rebuild toolchain; arm64 ABI alignment is verified pre-release via
-# `npm run local:gate` (see the inline note on the rebuild step below).
+# compliance verify, tests + coverage, IPC contracts, production webpack build) on
+# Ubuntu and Windows, and verify Electron-linked native modules load there.
 #
-# Push to long-lived branches (main/develop/RegressionFixes) runs the same gate to
-# catch regressions that land via non-PR pushes (e.g. hotfix commits direct-to-branch).
+# Push to long-lived branches (main/develop/RegressionFixes) runs the same gate on
+# all three OSes plus the E2E smoke. macOS and the E2E smoke are push-only to keep
+# Actions minutes down (macOS bills at ten times the Linux rate), so they report
+# after a merge, not before it: the ruleset "Main: required checks" requires only
+# `ci (ubuntu-latest)`, `ci (windows-latest)` and `CodeQL`. Run `npm run local:gate`
+# and the smoke specs locally before pushing. The macos-latest runner builds x64
+# while production ships arm64 — the x64 run still exercises the rebuild toolchain;
+# arm64 ABI alignment is verified pre-release via `npm run local:gate` (see the
+# inline note on the rebuild step below).
 
 name: CI
 
@@ -33,7 +37,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        os: ${{ fromJSON(github.event_name == 'pull_request' && '["ubuntu-latest","windows-latest"]' || '["ubuntu-latest","windows-latest","macos-latest"]') }}
     timeout-minutes: 60
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -99,13 +103,14 @@ jobs:
   # The full `npm run test:e2e` suite (including @live-llm specs) is run manually
   # before releases; there is no nightly E2E workflow.
   e2e-smoke:
+    if: github.event_name != 'pull_request'
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false
       matrix:
         os: [ubuntu-latest, windows-latest]
     # 45, not 20: each spec relaunches the app (~17 s/test on windows-latest),
-    # so the 91-test smoke list alone needs ~27 min plus ~5 min of npm ci +
+    # so the 101-test smoke list alone needs ~29 min plus ~5 min of npm ci +
     # native rebuild + dev builds before Playwright even starts.
     timeout-minutes: 45
     steps:
@@ -126,7 +131,7 @@ jobs:
         run: npm run build:dev
       - name: Run xvfb-friendly E2E smoke
         if: runner.os == 'Linux'
-        run: xvfb-run --auto-servernum -- npx playwright test test/e2e/app-launch.spec.js test/e2e/whisper-variant-loader.spec.js test/e2e/session-persistence.spec.js test/e2e/undo-redo.spec.js test/e2e/organize-phase.spec.js test/e2e/search-modal.spec.js test/e2e/settings-backup-restore.spec.js test/e2e/accessibility.spec.js --reporter=list
+        run: xvfb-run --auto-servernum -- npx playwright test test/e2e/app/app-launch.spec.js test/e2e/discover/whisper-variant-loader.spec.js test/e2e/app/session-persistence.spec.js test/e2e/organize/undo-redo.spec.js test/e2e/organize/organize-phase.spec.js test/e2e/search/search-modal.spec.js test/e2e/settings/settings-backup-restore.spec.js test/e2e/app/accessibility.spec.js --reporter=list
         env:
           CI: 'true'
           # pw:browser* forwards the Electron child's stdout/stderr from spawn
@@ -135,7 +140,7 @@ jobs:
           DEBUG: 'pw:browser*'
       - name: Run E2E smoke (Windows)
         if: runner.os == 'Windows'
-        run: npx playwright test test/e2e/app-launch.spec.js test/e2e/whisper-variant-loader.spec.js test/e2e/session-persistence.spec.js test/e2e/undo-redo.spec.js test/e2e/organize-phase.spec.js test/e2e/search-modal.spec.js test/e2e/settings-backup-restore.spec.js test/e2e/accessibility.spec.js --reporter=list
+        run: npx playwright test test/e2e/app/app-launch.spec.js test/e2e/discover/whisper-variant-loader.spec.js test/e2e/app/session-persistence.spec.js test/e2e/organize/undo-redo.spec.js test/e2e/organize/organize-phase.spec.js test/e2e/search/search-modal.spec.js test/e2e/settings/settings-backup-restore.spec.js test/e2e/app/accessibility.spec.js --reporter=list
         env:
           CI: 'true'
       - name: Upload E2E artifacts on failure

CHANGELOG.md

@@ -12,6 +12,20 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Fixed
+
+- **An image that arrives just as a text model starts loading no longer starts the vision runtime
+  beside it.** The image looked for a loading model before the load began and found none. The vision
+  runtime now checks the memory budget again at the moment it starts; if a load began meanwhile, the
+  image waits for it, makes room and is sent once more. Such a refusal does not count toward turning
+  vision off.
+- **An embedding model put on the CPU beside a text model returns to the GPU when there is room.**
+  After a text-model load that failed, or once the text model was unloaded, it stayed on the CPU
+  until some other load replaced it.
+- **A text model larger than the memory budget no longer unloads models that hold no GPU memory.**
+  An embedding model on the CPU and Whisper on the CPU were unloaded to make room, which freed
+  nothing, and were loaded again for the next file.
+
 ## [5.2.0] - 2026-11-10
 
 Cloud analysis now covers audio and video, and video frames are sent only with your consent and the

CLAUDE.md

@@ -60,7 +60,7 @@ npm run test:e2e         # Playwright E2E tests
 npm run lint             # ESLint check
 npm run format:check     # Prettier check
 npm run ci               # format:check + lint + typecheck + audit:production + compliance:commercial:verify + test:coverage + verify:ipc-contracts + build, plus the full verify:* suite (channels, ipc handlers/schemas/enforcement, mcp + ai-provider contracts, analysis routing, electron fuses, env docs, version tag) — package.json is the authoritative list
-npm run local:gate       # verify:installed-electron + ci + verify:native-modules (one-shot pre-push convenience). GitHub Actions `CI` is configured for Windows + Ubuntu + macOS when Actions are enabled for the repo. Cloud CI has been off at the org/repo setting at times — `dependabot.yml` notes local gate as the only guaranteed gate. The e2e-smoke job is restricted to Windows + Ubuntu.
+npm run local:gate       # verify:installed-electron + ci + verify:native-modules (one-shot pre-push convenience). GitHub Actions `CI` runs Ubuntu + Windows on pull requests and adds macOS + the e2e-smoke job on pushes to main (trimmed 2026-11-11 to cut Actions minutes; required checks are `ci (ubuntu-latest)`, `ci (windows-latest)` and `CodeQL`), when Actions are enabled for the repo. Cloud CI has been off at the org/repo setting at times — `dependabot.yml` notes local gate as the only guaranteed gate. The e2e-smoke job is restricted to Windows + Ubuntu.
 npm run dist:win         # Build Windows installer
 ```
 
@@ -188,8 +188,10 @@ These are tracked gaps that the project slash commands help address:
   counts are the output of [`.github/workflows/ci.yml`](.github/workflows/ci.yml) — treat CI logs as
   source of truth rather than inline numbers here.
 - ~~Coverage not enforced in CI~~ - CI now runs `npm run test:coverage` (enforces thresholds).
-  Cross-platform matrix (Windows, macOS, Linux). Coverage artifacts uploaded.
-- ~~Single-platform CI~~ - CI runs on `windows-latest`, `ubuntu-latest`, and `macos-latest`.
+  Cross-platform matrix (Windows and Linux on pull requests; macOS added on pushes to main).
+  Coverage artifacts uploaded.
+- ~~Single-platform CI~~ - CI runs on `windows-latest` and `ubuntu-latest` for pull requests, plus
+  `macos-latest` on pushes to main.
 - ~~No IPC handler coverage check~~ - `npm run verify:ipc-handlers` statically verifies 100% of
   `IPC_CHANNELS` have registered handlers. `npm run ci` (run by every CI job) transitively invokes
   `npm run verify:ipc-contracts` — there is no standalone `ipc-contract-check` job.

docs/TESTING.md

@@ -290,6 +290,10 @@ Notes:
 
 E2E automation notes:
 
+- Specs are grouped by product area under `test/e2e/<area>/` (`app`, `setup`, `discover`,
+  `organize`, `search`, `settings`, `workflows`, `live-llm`, `manual`).
+  [`test/e2e/README.md`](../test/e2e/README.md) lists every spec, the CI smoke set and what each
+  spec needs. Run one area with `npx playwright test test/e2e/<area>`.
 - Every `launchApp()` profile is isolated and seeded with `modelAutoDownloadSkipped: true`
   (`E2E_SEED_SETTINGS` in `test/e2e/helpers/electronApp.js`), so a run never starts a model download
   and needs no network. A spec that needs the first-run model wizard keeps the seed and calls

docs/commercial-use-inventory.json

@@ -1,5 +1,5 @@
 {
-  "generatedAt": "2026-11-10T23:39:10.960Z",
+  "generatedAt": "2026-11-12T01:40:35.862Z",
   "projectLicense": {
     "riskLevel": "allowed",
     "reason": "Project license is PolyForm Noncommercial 1.0.0 for personal/noncommercial use. Commercial rights are sold separately; third-party dependency obligations remain governed by their own licenses.",
@@ -18725,7 +18725,7 @@
     },
     {
       "packageName": "officeparser",
-      "version": "8.6.0",
+      "version": "8.7.0",
       "license": "MIT",
       "riskLevel": "allowed",
       "directRuntime": true,

package.json

@@ -104,7 +104,7 @@
     "tour:record:studio": "cross-env STRATOSORT_REDACT_PATHS=1 node scripts/record-product-tour.js --topic=studio",
     "tour:compose": "node scripts/compose-walkthrough-media.js",
     "tour:compose:lessons": "node scripts/compose-walkthrough-media.js --topic=lessons",
-    "test:soak": "cross-env SOAK=1 SOAK_MINUTES=12 npx playwright test test/e2e/soak.spec.js --timeout=1200000",
+    "test:soak": "cross-env SOAK=1 SOAK_MINUTES=12 npx playwright test test/e2e/manual/soak.spec.js --timeout=1200000",
     "models:check": "node scripts/setup-models.js --check",
     "models:download": "node scripts/setup-models.js --download",
     "dist": "npm run build && npm run generate:nsis && npm run setup:vision-runtime && npm run setup:vision-runtime:check && npm run setup:ffmpeg-runtime && npm run setup:ffmpeg-runtime:check && npm run setup:tessdata-runtime && npm run setup:tessdata-runtime:check && npm run setup:vc-redist -- --windows-host-only && npm run setup:vc-redist:check -- --windows-host-only && electron-builder --publish never --config electron-builder.json",
@@ -228,7 +228,7 @@
     "mammoth": "^1.13.0",
     "minimatch": "^11.0.5",
     "node-llama-cpp": "^3.22.3",
-    "officeparser": "^8.6.0",
+    "officeparser": "^8.7.0",
     "p-queue": "^10.0.0",
     "pdfjs-dist": "6.5.438",
     "pino": "^10.4.0",

playwright.config.js

@@ -19,10 +19,11 @@ const { defineConfig } = require('@playwright/test');
 const path = require('path');
 
 module.exports = defineConfig({
-  // Directory containing E2E test files
+  // Directory containing E2E test files. Specs are grouped by product area in
+  // subfolders (app/, setup/, discover/, ...); see test/e2e/README.md for the index.
   testDir: './test/e2e',
 
-  // Test file pattern
+  // Test file pattern (recursive, so every area folder is picked up)
   testMatch: '**/*.spec.js',
 
   // Maximum time one test can run (3 minutes for analysis tests)

src/main/ipc/files/fileOperationHandlers.js

@@ -249,6 +249,28 @@ function getAnalysisHistoryService() {
   return null;
 }
 
+/**
+ * Stored analysis of a file that was just moved, so learning can key the move
+ * on the file's category. The history entry follows the file to `destPath`;
+ * `sourcePath` covers a path update that did not land.
+ * @param {string} sourcePath
+ * @param {string} destPath
+ * @returns {Promise<Record<string, any>|null>} Analysis, or null when the file was never analyzed
+ */
+async function getMovedFileAnalysis(sourcePath, destPath) {
+  const historyService = getAnalysisHistoryService();
+  if (!historyService?.getAnalysisByPath) return null;
+  try {
+    const entry =
+      (await historyService.getAnalysisByPath(destPath)) ||
+      (await historyService.getAnalysisByPath(sourcePath));
+    return entry?.analysis || null;
+  } catch (error) {
+    logger.debug('[FILE-OPS] Analysis lookup for learning failed:', error?.message);
+    return null;
+  }
+}
+
 /**
  * @param {{ sourcePath: string|null|undefined, destPath: string, log: import('../ipcWrappers').LoggerLike, context: string }} args
  */
@@ -700,7 +722,9 @@ async function moveOneFile(
       await learningService.recordFileMove(
         moveValidation.source,
         moveValidation.destination,
-        null, // No analysis available for manual moves
+        // Without the category the pattern is `<ext>:unknown:<folder>`, which
+        // an analyzed file's lookup never matches.
+        await getMovedFileAnalysis(moveValidation.source, moveValidation.destination),
         FEEDBACK_SOURCES.MANUAL_MOVE
       );
     }