Release readiness: docs accuracy, test coverage, and UI code-behind decomposition

[DevMando]

Summary

Pre-release hardening of MandoCode Desktop. This branch makes the documentation match the shipped app, adds automated tests around the highest-risk logic, and decomposes the two large UI code-behind files into focused, maintainable units — with no changes to the shared MandoCode engine and no behavioral changes to the app.

Note on scope: this branch also carries the previously-built session history / compare view / snapshot feature commits, which had not yet reached main. Those are the two feature commits below the hardening commit; the sections here describe the hardening work, which is the primary review target.

Why now

The app was functionally solid but carried three sources of release risk:

1. Documentation had drifted from the code — including a changelog that described a data model we had removed, and several shipped features that were undocumented.
2. Test coverage was thin and concentrated on low-risk helpers, with nothing covering the code most likely to break.
3. Two UI files had grown to ~2,800 lines each, making them slow to navigate and prone to merge conflicts.

This pass pays that down so we can tag a release with confidence.

What changed

Documentation accuracy

• Reconciled CHANGELOG.md / README.md with the shipped app: removed descriptions of a snapshot data model that no longer exists, and documented three features that shipped but were undocumented (integrated terminal, file-explorer + git awareness, AI skill authoring).
• Fixed a developer-facing build-guard error message that pointed at APIs which don't exist.

Safety net — automated tests raised from 40 to 73

• Config clone + validate round-trip — covers a case-sensitivity bug in MCP-server lookups that the code repeatedly warned about but never tested.
• Agent-name slot reuse, conversation-history flattening, and the invisible request-preamble that gets folded into each model message.
• The streamed-response loop — the highest-risk path in the app — is now covered, driven by a fake AI-service seam so it runs without a live model (empty response, cancellation, error, and cloud-auth cases are all asserted).

Maintainability — behavior-preserving refactors

• Decomposed MainWindow (2,877 → 152-line core + 12 focused files) and ChatTabView (2,414 → 351 + 7 files) via partial-class splits. Same classes, same behavior, far more navigable.
• Moved the transcript view's ~970 lines of embedded CSS/JS out of C# and into asset files (1,258 → 344).
• Introduced a small IAiService interface so the UI controller depends on an abstraction rather than the concrete engine — this insulates it from engine API changes and is what makes the streaming loop testable.
• Removed duplicated helpers (project/label formatting, shell-open, config cloning) into single sources of truth.

Bug fixes

• The MCP server-editor modal rendered pinned in the left sidebar rail instead of centered over the content (a latent layout bug).
• Tightened the chat input placeholder so it stays legible in the side-by-side compare view.

Risk & verification

• The shared MandoCode engine (git submodule) is untouched and its pinned version is unchanged, so rolling it forward later works exactly as before.
• Refactors are behavior-preserving (partial-class splits and pure extractions); the transcript HTML output is unchanged.
• Build is clean (0 warnings, 0 errors) and all 73 tests pass.
• Manually smoke-tested: chat streaming, all approval types (write/command/delete/MCP), request cancellation, model switching, theme switching (including the e-ink/CRT/Win98 modes), chat background, tabs, split/compare view, the integrated terminal, and each settings page.

Deferred (non-blocking, tracked for after release)

• Deeper decoupling of the file-explorer code-behind and the MCP/snapshot command handlers.
• Extracting the approval overlay and the snapshots/history panels into standalone controls — the partial-class splits already resolve the file-size concern, so these are polish rather than blockers.