Complete MCP Read Evidence Ladder

The deployed read/evidence substrate improves MCP and CLI structure, but hosted ChatGPT retesting still found dead ends: search handles without matched text, field-window resource URIs that the client could not read, and full fetches that materialized files for small text records.

MCP clients differ in what they expose to the model. Some expose structuredContent, some surface only content[], some support resources, and some show resource links but cannot read them. The current substrate assumes too much from the client: it may put enough information in structured fields or resource handles while visible text remains insufficient for classification.

The MCP adapter SHALL render search results so a client that only exposes content[] can decide whether to inspect a result further without reading structuredContent.

The resource server SHALL expose bounded search match-window evidence when it can identify the text field and character window that caused a search hit. The resource server SHALL NOT fabricate a match window when the backend cannot prove the matched field.