chat: use structured outputs for SearchBookContent MCP tool

[BenjaminMichaelis]

Summary

Improves structured output consistency across the AI chat RAG pipeline by converting SearchBookContent — the primary semantic vector search tool — to use the same hybrid structured content pattern already established by FindBookHelpForDiagnostic and SearchListingsByCode.

The changes were reviewed and approved by Claude Opus 4.6 and GPT-5.5 subagents, iterating to agreement.

Changes

SearchBookContent converted to structured output

SearchBookContent was the only major RAG tool still returning plain text. It now returns a CallToolResult with:

• UseStructuredContent = true and OutputSchemaType = typeof(SearchBookContentResult)
• McpToolResultFormatter.GetModelInput() will now return the JSON StructuredContent to the model instead of the formatted text, consistent with all other structured tools
• The human-readable text is preserved as the first Content block for MCP clients that display it

"No results" semantics fixed

The zero-results path previously returned McpToolResultFactory.CreateError() (IsError = true). Zero results from a search query is a valid outcome, not a tool failure. It now returns CreateHybridResult with an empty Matches: [] list, so the model always receives well-typed JSON conforming to the schema.

Validation failures (empty query, query too long, AI service not configured) remain correctly modeled as errors.

New structured result types

Added to McpToolResults.cs:

• SearchBookContentMatchResult — one vector search hit (score, chapter number, heading, text chunk)
• SearchBookContentResult — wrapper with Matches list; this is the advertised output schema

OutputSchemaType made explicit on all UseStructuredContent tools

Five tools that had UseStructuredContent = true but no OutputSchemaType were updated:
GetChapterList, GetChapterSections, GetDirectContentUrl, GetNavigationContext, GetChapterSummary.

The MCP SDK already auto-infers the output schema from the concrete return type (confirmed by pre-existing integration tests), so this is redundant but makes the intended schema contract explicit at the declaration site — consistent with how FindBookHelpForDiagnostic and SearchListingsByCode are declared.

Contract test updated

McpToolContractTests.McpToolsList_StructuredAndHybridTools_AdvertiseOutputSchema now asserts that search_book_content advertises an outputSchema, alongside the other structured tools.

What was NOT changed (and why)

• Main chat response format — ResponseTextFormat.CreateJsonSchemaFormat() does not exist in OpenAI SDK 2.7.0 for the Responses API. Schema-enforced output on the streaming text response isn't possible without switching to the Chat Completions API, and would break markdown rendering anyway. Both subagents agreed this is correct.
• strictModeEnabled: true on tool input registration — Already correct in AIChatService.cs, untouched.
• Plain-text tools (LookupConcept, CheckTopicCoverage, FindRelatedSections, etc.) — These return prose-heavy navigation summaries where structured content wouldn't add value for the model.

Testing

• ✅ Solution builds: 0 errors, 0 warnings
• ✅ All 11 chat unit tests pass
• ✅ Contract test updated to cover search_book_content schema advertisement