Fix: keep MCP diagnostics observations within the model context budget

[tninja]

Summary

The diagnostics_baseline MCP tool was returning the full project diagnostics list in its response. On a checkdoc-heavy Elisp project that payload reached ~245 KB and could not fit in the model context, which broke the diagnostics-first TDD harness (the harness calls diagnostics_baseline before editing). Since the baseline is meant to be recorded server-side, echoing every diagnostic back into the model context contradicted its own design. This PR makes the diagnostics observations context-safe and a bit more actionable.

What changed

• diagnostics_baseline no longer echoes the diagnostics list. The baseline is already stored server-side in the identity-count table, so the response now carries only status, a summary (with the count), and a structured next_actions step. The payload drops from ~245 KB to a constant ~120 B regardless of project size.
• get_diagnostics reports are bounded. The shared observation envelope now caps files and next_actions at a new defcustom ai-code-mcp-diagnostics-max-report-diagnostics (default 200). The summary always reports the true totals and notes when the list was truncated, so the project-wide path cannot overflow the context either.
• More signal, less bulk. The baseline summary gains a Top sources: breakdown (which checkers dominate), and the structured next_actions tells the agent exactly how to verify — edit, then call get_diagnostics with since="baseline" until status is clean.

Verification

• New ERT tests cover each behavior: the baseline response omits the list, the envelope truncates large reports, the summary names top sources, and the baseline response carries next_actions. The full test_ai-code-mcp-server.el suite passes (38/38) and the module byte-compiles with no warnings.
• Verified live in Emacs: the baseline payload is constant (~120 B vs ~245 KB before), and the end-to-end harness loop converges correctly — baseline then regression when a new diagnostic appears, then clean once it is fixed.

[Copilot]

Pull request overview

This PR reduces MCP diagnostics payload size so diagnostics-first harnesses can record baselines and inspect reports without overflowing model context.

Changes:

• Adds a configurable cap for diagnostics included in get_diagnostics observation envelopes.
• Changes diagnostics_baseline to store diagnostics server-side without echoing the full file list.
• Adds tests for baseline payload shape, truncation, source summaries, and structured next actions.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
ai-code-mcp-server.el	Implements capped diagnostics envelopes and compact baseline responses.
test/test_ai-code-mcp-server.el	Adds ERT coverage for the new diagnostics response behavior.

[tninja]

feedbacks have been addressed.