[iLevyTate]

What changed

• Atlas: "Smart folder…" acts on the current selection.
• Chat: Regenerate and Retry remove the discarded answer from saved history, so a reload no longer shows both answers.
• Learning: manual moves teach, undo retracts the matching pattern, and auto-organize records feedback once, after the move succeeds (it used to record at plan time and again after the move).
• Search index: a warning appears when saving the index keeps failing, and clears on the next successful save.
• Small GPUs: vision runtime memory is checked and reserved in one step, the embedding model returns to the GPU after vision work, and zero-byte CPU-beside residents no longer count against the budget.
• Dead route: the unused watcher reanalyze channel, handler and schema are removed.
• Dependency: officeparser 8.6.0 → 8.7.0 (supersedes build(deps): Bump officeparser from 8.6.0 to 8.7.0 in the minor-and-patch group #119).
• CI: pull requests run ci on Ubuntu and Windows; macOS and e2e-smoke run on push only.
• E2E: specs are grouped into feature folders. Playwright lists the same 387 tests in 40 files before and after.

Verification (local, Windows x64)

• npm run ci steps: format, lint, typecheck, audits, compliance, test:coverage (947 suites, 13,973 tests passed, 4 skipped, thresholds met), all verify:* steps, production build.
• npm run verify:native-modules: 8 checks pass.
• The eight smoke specs on a clean build:dev: 101 tests pass.

Not verified

• The small-GPU fixes have unit tests only; they have not been confirmed on a real small-GPU machine or Apple Silicon.
• Nothing was run on macOS or Linux locally.

Left open

• Undoing a single manual move does not retract what it taught.
• Download-watcher moves teach nothing; the smart-folder watcher can re-learn after re-analysis.
• The index-save warning has no idle retry; it retries when the index next changes.