Build the Scantle control plane, isolated candidate verifier, contract diff engine, and orders reference platform

.gitignore

@@ -0,0 +1 @@
+ui/checkpoint/node_modules/

PRESS_RELEASE.md

@@ -10,7 +10,7 @@
 
 Engineering leaders adopting AI coding agents face a growing maintainability and reliability gap. Empirical research from **GitClear** (*"Coding on Copilot: 2023 Data Suggests Downward Pressure on Code Quality,"* 2024; *"AI Copilot Code Quality: Evaluating 2024's Increased Defect Rate via Code Quality Metrics,"* Feb 2025, analyzing 211 million changed lines of code; and [*"The Maintainability Gap: AI Code Quality in 2026"*](https://www.gitclear.com/the_ai_code_quality_maintainability_gap), analyzing 623 million changes across 2023–2026) observed declining code reuse/refactoring, rising two-week code churn, and a **+81% increase in its 2023–2026 duplication measure**, while **Tornhill & Borg’s *Code Red* study (2022)** across 39 proprietary production codebases demonstrated that low-health code contains **15× more defects** and takes **124% longer** to resolve issues than healthy code. When teams build distributed agentic platforms—which require typed tool schemas, MCP servers, durable execution across crashes and multi-day human approvals, live event streaming, and observability—either humans spend weeks wiring infrastructure glue by hand, or autonomous "dark factories" burn token budgets maintaining repetitive boilerplate and weakening their own tests to pass CI.
 
-Scantle addresses these challenges by pairing **Model-Driven Software Product Line code generation**—motivated by **CMU Software Engineering Institute (SEI)** product-line case studies showing reduced ongoing maintenance effort and faster time-to-market—with **process-isolated behavioral verification**. Instead of asking LLMs to write and maintain transport and schema plumbing, Scantle expresses services, agent toolsets, MCP endpoints, evaluation suites, and C4 architecture diagrams in a single executable Go design (`design/*.go`). Every time a capability changes, `goa gen` deterministically compiles the `gen/` transport, schema, and workflow layer at **zero LLM token cost** (in `chorus`, for example, 239 lines of Goa design generate 21,686 lines of `gen/` code alongside 2,566 lines of handwritten Go). Builder agents receive only the compact design slice and the exact Go interface signatures they need to implement, coordinating in real time over **Chorus** with zero-token background waiting (`chorus watch`). For production runtime efficiency, Scantle scaffolds **Goa-AI’s runtime token and cost controls** by default—including application-supplied input injection (`Inject` and direct `BindField`), bounded tool views with runtime-held pagination cursors (`BoundedResult`/`ContinueWith`), UI-only data streams (`ServerData`), isolated child-agent workflows (`Use(AgentToolset(...))`), and provider-supported prompt-cache breakpoints (`RunPolicy` `Cache`, which unlock 75%–90% cached-input price discounts on providers that support explicit checkpoints such as Anthropic)—and enforces **per-scenario runtime token and cost budgets** in `scantle.verify` on every change.
+Scantle addresses these challenges by pairing **Model-Driven Software Product Line code generation**—motivated by **CMU Software Engineering Institute (SEI)** product-line case studies showing reduced ongoing maintenance effort and faster time-to-market—with **process-isolated behavioral verification**. Instead of asking LLMs to write and maintain transport and schema plumbing, Scantle expresses services, agent toolsets, MCP endpoints, evaluation suites, and C4 architecture diagrams in a single executable Go design (`design/*.go`), guided by Scantle’s built-in **Golden Path Architecture Playbook** (`scantle://golden-path`). Every time a capability changes, `goa gen` deterministically compiles the `gen/` transport, schema, and workflow layer at **zero LLM token cost** (in `chorus`, for example, 239 lines of Goa design generate 21,686 lines of `gen/` code alongside 2,566 lines of handwritten Go). Builder agents receive only the compact design slice, the generated `AGENTS_QUICKSTART.md` agent/tool contract guide, and the exact Go interface signatures they need to implement, coordinating in real time over **Chorus** with zero-token background waiting (`chorus watch`) and deterministic **cross-family role routing** (`@architect`, `@builder-<scope>`, `@eval-author`, and `@reviewer`) that pairs builder models with a disjoint model family for evaluation authoring and adversarial review whenever eligible capabilities and scope assignments permit. For production runtime efficiency, Scantle scaffolds **Goa-AI’s runtime token and cost controls** by default—including a central stateless gRPC inference gateway (`goa-ai/features/model/gateway` with `middleware.NewAdaptiveRateLimiter` and `pulse` capacity coordination), application-supplied input injection (`Inject` and direct `BindField`), bounded tool views with runtime-held pagination cursors (`BoundedResult`/`ContinueWith`), UI-only data streams (`ServerData`), isolated child-agent workflows (`Use(AgentToolset(...))`), and provider-supported prompt-cache breakpoints (`RunPolicy` `Cache`, which unlock 75%–90% cached-input price discounts on providers that support explicit checkpoints such as Anthropic)—and enforces **per-scenario runtime token and cost budgets** in `scantle.verify` on every change.
 
 > *[Draft Spokesperson Quote — For Approval by Raphael Simon]*
 > "Unconstrained coding agents act like developers perpetually new to your codebase—they copy-paste transport boilerplate, burn tokens re-reading entire repositories, and grade their own homework," said **Raphael Simon, creator of Goa and the Goa Design suite**. "Scantle turns your favorite AI assistant into a disciplined software factory over MCP. Goa and Goa-AI compile your contracts deterministically so your team maintains a fraction of the code and coding agents spend tokens only on business logic—while an isolated evaluator blocks any change that breaks crash recovery or causes a workflow’s runtime token cost to regress."