feat: add remote-first education work lake

This commit is contained in:
冰朔 2026-08-07 20:41:03 +08:00
commit 333cd222c5
43 changed files with 923 additions and 74 deletions

View file

@ -0,0 +1,27 @@
---
type: ADR
id: "0173"
title: "Remote-first industry work lakes use registered client-rendered modules"
status: active
date: 2026-08-07
---
## Context
The inherited Tolaria entry assumes that a local filesystem vault must exist before any useful surface can open. HoloLake needs a different product boundary: Git repositories hold reusable, numbered components; a persona routes human language to those components; the resulting document, spreadsheet, chart, dashboard, or knowledge surface renders on the user's device. An industry workspace may be temporarily hosted on `JD-FD-PRIMARY` and later moved without changing its capability identifiers.
## Decision
Industry work lakes are remote-first application channels above the optional local vault.
- The HoloLake world entry is available without creating, choosing, or repairing a local vault.
- Every callable surface has a stable module id in `engineering/MODULE-REGISTRY.json`.
- Language routing selects a registered capability and emits a receipt; it does not treat model text as proof of execution.
- Stage-one document and spreadsheet state remains inside the application/browser boundary.
- Server hosting and client rendering are separate: the temporary server serves the application, while user-facing editing and projection execute on the client.
- The first education work lake mounts document, spreadsheet, comparison-chart, dashboard, and knowledge surfaces. Remote multi-user persistence is a later capability and must have its own authorization and receipt contract.
- Grok Build remains an isolated research source for agent-runtime patterns. Its runtime is not merged into a persona subject.
## Consequences
The obsolete local-vault chooser no longer blocks the HoloLake entry. Existing vault features remain available as one optional knowledge module. Industry migration becomes a deployment-route change rather than a data-model rewrite. Collaborative persistence, domestic model providers, and production whiteboards remain explicit follow-up modules instead of hidden claims in stage one.