Feature: show how full each agent's context window is - #79

Summary

Each tab now shows how full its agent's context window is, in the strip under the message box:

Context ▰▰▰▰▰▱▱▱▱▱▱▱ 41% of 32k tokens              ⎇ ● main

The bar and the percentage turn from green to yellow to red as the window fills. That's the cue to run /compact before the model starts losing the earlier part of the conversation, which small local models do quickly.

What changed

• Meter in the status strip: it sits left of the git branch chip, and every split pane has its own, since context is per agent. The strip now shows whenever either item has something to display; before, it showed only for git repositories.
• What it measures: the size of the last request against the model's configured context window. Every request resends the whole conversation, so that's where the next one starts. It turns yellow at 60% and red at 85%, and a full window also shows ⚠ so the warning doesn't rely on color alone.
• Cloud models: these manage their window server-side, so they show the token count only (~17k tokens in context), without a bar.
• Details on hover: the tooltip gives the exact count ("13k of 32k tokens") and suggests /compact. Screen readers announce it as "Context window 41 percent full, 13k of 32k tokens".
• Colors checked in every theme: each theme's green, gold and red are lifted just enough to stay readable. The percentage, which is text, clears 4.5:1 against the background. The bar clears 3:1 against its empty track, so the filled part is always easy to tell apart. Measured before this change, the raw colors missed even 3:1 in W98 - Y2K and Solarized Light, and the filled part would have been nearly the same color as the empty part in most themes.

Why this was done

The only context signal Desktop had was the per-reply summary line ([~27k in, 611 out]) and a session-wide total in the header. Neither says how close the agent is to running out of room. The CLI gained the same meter in DevMando/MandoCode#101, and this brings Desktop in line.

How it was tested

• A new theme test checks every theme at every level (25 themes × 3 levels), so a future theme with a weak color fails the test instead of shipping. The full Desktop test suite passes.
• Worth a manual check in the app before merging: send a message on a local model and confirm the meter appears after the reply, then switch themes and confirm it recolors.

Notes

• While a plan runs, the meter reflects the last plan step's request, which is smaller than the chat's own context. The CLI meter behaves the same way.
• The math comes from the engine's ContextMeter, already on the pinned engine commit; this PR doesn't move the pin.
• Changelog: entry added under [Unreleased] → Added.