Expose bounded AST diff summaries in MCP results

[thromel]

Why

The MCP tools can already report detected refactorings, but that is often not enough for an agent reviewing a change. For many refactorings, the Git diff is the wrong first view: moved or extracted code turns into a large delete/add block, and the agent has to read a lot before it knows where the structural edits are.

This PR adds a bounded astDiffs field to MCP results. It gives clients a small AST-level index of the change: file pairs, mapping/action counts, counts by action type, sampled actions, line ranges, update values, and move metadata.

Changes

• Added McpAstDiffResult.
• Return astDiffs from analysis tools for file contents, worktrees, commits, pull requests, and directories.
• Added maxAstDiffs and maxActionsPerAstDiff; maxAstDiffs=0 omits the payload.
• Return the same bounded summaries from WebDiff tools while preserving localhost URL behavior.
• Tightened sampled action details:
  • missing positions use -1 for offsets and lines
  • GumTree end offsets are treated as exclusive
  • multi-move actions use a real targetFilePath plus separate moveGroupId
• Reject fractional numeric MCP arguments instead of truncating them.
• Updated documentation/mcp.md.

What this gives MCP clients

A client can use RefactoringMiner as the first pass in review: see which file pairs changed structurally, which action types dominate, and which nodes deserve source inspection. It does not try to prove a refactoring is safe. It gives the agent a better map before it opens files.

Validation

• ./gradlew test --tests org.refactoringminer.mcp.RefactoringMinerMcpServiceTest --tests org.refactoringminer.mcp.RefactoringMinerMcpToolsTest --tests org.refactoringminer.mcp.WebDiffBrowserLauncherTest --no-daemon --console=plain
• ./gradlew shadowJar mcpSmoke --no-daemon --console=plain
• git diff --check
• Live Claude Code MCP checks with Sonnet 4.6:
  • file-content analysis returns bounded astDiffs
  • worktree analysis uses astDiffs on this RefactoringMiner branch
  • all 14 MCP tools were exercised in a real tool sweep
  • WebDiff tools return URLs while respecting maxAstDiffs=0
  • fractional maxRefactorings=1.5 is rejected with an error

Related to #1058.