perf(chat): cache context bar rendering

[itkonen]

I noticed the CPU load increased after the introduction of the context bar - enough to make it the bottleneck on older computers - so I did some benchmarking and found the problem. I'm not quite sure if this fix is ideal, but at least it brought down the CPU load.

AI Summary

• Cache the context usage mode-line bar so streaming redisplay does not rebuild the same propertized bar and tooltip on every mode-line evaluation.
• Use a normalized cache key based on the actual rendering inputs, including context totals/categories, graphical state, frame char width, and fallback colors.
• Add a focused streaming benchmark entry point for comparing default, no-context, no-mode-line, mode-line, context-bar, and usage-event paths.

Verification

• emacs -Q --batch -L . ... --batch -f batch-byte-compile eca-chat.el benchmarks/eca-chat-bench.el
  • passes; only pre-existing optional Evil warnings remain.
• emacs -Q --batch -L . -L test ... -l buttercup -l test/eca-chat-test.el -f buttercup-run
  • 79 specs, 0 failed.
• emacs -Q --batch -L . ... -l benchmarks/eca-chat-bench.el -f eca-chat-bench-run-streaming
  • confirms cached context-bar mode-line rendering remains close to the no-context path.
• git diff --check
  • clean.

🤖 Generated with ECA (openai/gpt-5.5 - high)