ADR 0321: Diagram geometry follows Markdown code-block semantics

Status: accepted; amends the geometry contract of ADR 0187 without changing the freeform opt-out.

discern tidy validates box-drawing alignment in configured Markdown because a formatter can preserve syntactically valid text while leaving a diagram visually broken. The original scanner enrolled fenced code blocks. Markdown also defines indented code blocks, but box-drawing content there bypassed the geometry check. A misaligned diagram could therefore fail in fenced form and pass unchanged when represented with four-space indentation.

The distinction is syntax, not intent. Both block forms represent code-like literal content and preserve the same box-drawing geometry. An agent should not need to know that the validator's parser happened to recognize only one spelling.

Fenced blocks have an info string, which supports the existing freeform opt-out for content whose columns are not structural. Indented blocks have no equivalent metadata channel.

Box-drawing geometry validation follows Markdown code-block semantics, covering both fenced and indented blocks.

• The Markdown block scanner emits literal content for fenced and indented code blocks through one diagram validator. A box-drawing glyph in either form enrolls the block.
• Both forms use the same display-column alignment rule and the same diagrams_misaligned diagnostic with file, line, and column.
• The existing fenced freeform info-string tag remains the explicit opt-out. Because an indented block carries no info string, freeform box-drawing content must use a fenced block tagged freeform.
• Tidy's all-or-nothing planning remains unchanged: any misaligned diagram stops the run before a Markdown or TOML target is written.
• Fixtures cover aligned and misaligned examples in both forms and prove the freeform exemption remains limited to the tagged fenced form.

• Changing a diagram from fenced to indented syntax cannot bypass the Gate's geometry contract.
• Diagnostics and recovery remain identical across block forms; authors fix alignment or choose the visible freeform spelling.
• The scanner must implement Markdown indentation and blank-line boundaries accurately enough to identify the literal block without treating ordinary nested prose as code.
• A future Markdown literal-block form must join the same semantic scanner or carry a reasoned exception; syntax-specific enrollment is no longer sufficient.

• Normalize every indented block to a fence. Rejected because tidy promises canonical formatting without rewriting literal block style or inventing info strings.
• Leave indented diagrams unchecked. Rejected because the same visual defect would keep two verdicts based only on Markdown spelling.
• Treat every indented block as freeform. Rejected because that preserves the bypass and makes the opt-out implicit.
• Invent an adjacent comment opt-out. Rejected because the existing fenced freeform tag is visible, local, and sufficient; a second mechanism would create competing authority.