Context window setting is now live, per agent tab

[DevMando]

Summary

Companion to the MandoCode engine PR "Send the context window with every request and compact before overflow" (DevMando/MandoCode). It picks up the engine change that makes the context window setting real in every setup, and updates the desktop app's settings screen and messages that still described the old behavior. Merge the engine PR first, then this one.

What changed

• Engine update (submodule pin). The context window now travels with every chat request instead of being applied only when MandoCode starts the Ollama engine itself — so it works even when the Ollama desktop app owns the engine — and conversations automatically compact before they outgrow the window instead of failing with an empty response.
• The setting is now truly per-agent. Each agent tab's context window governs that tab's own conversations — something the old one-engine-one-window design couldn't do. The Settings field no longer carries the "app-wide" label, and its caption now explains that changes apply from the agent's next message with no restart.
• Messages match reality. The model-switch notice used to say the new window size "takes effect when MandoCode starts Ollama"; it now says it applies from the next message. Internal architecture notes (README, ConfigCoordinator) and the CHANGELOG are updated to match.

Why it matters

The desktop settings screen was promising something it couldn't deliver: users could set a context window, see a confirmation, and still be running on whatever the Ollama desktop app happened to be set to. That mismatch was the root cause of small local models failing mid-conversation with an empty response.

Scope and risk

• UI changes are text-only (a label, a caption, one transcript message) — no layout or behavior changes in the desktop code itself.
• The behavior change arrives via the engine pin, which is covered by its own PR and test suite.

Verification

• 223/223 desktop unit tests pass; the app builds clean against the updated engine.
• Verified on hardware: model switches resize the live window automatically (8k on a sub-3B model, 32k on a 9B model), no restarts involved.