refactor(ghost_text): handle multibyte characters and reduce render calls

[soifou]

Fixes ghost text display with multibyte characters showing as gibberish. Was using byte offsets instead of character counts, now properly uses strchars() and strcharpart() for UTF-8 handling.

Demo

Before:

2025-11-14_12-25-59.mp4

After:

2025-11-14_12-27-10.mp4

Also addresses a few fixes for noice.nvim ; the ghost text did not appear in cmdline for some context, e.g. :cd<Space> and an extra redraw was making the search highlighted text blinking.

Refactored to keep only the event-driven architecture:

• Removed decoration provider in favor of explicit list events
• Pass context object to avoid repeated mode/line lookups
• Drop TextChangedI but retained CursorMovedI autocmd to prevent text shaking during typing.

Overall this translate in a small performance improvements. The draw calls are reduced by 22-29% and total CPU time reduced by 9-95% depending on operation.

Performance benchmarks details

Typing vim.fn.getcmdtype with slow typing, fast typing, and extended 500-iteration test.

Draw Preview Performance

Branch	Test Type	Calls	Total Time	Avg per Call	vs Main
Main	Slow	56	11.29ms	0.20ms	baseline
Main	Fast	45	5.79ms	0.13ms	baseline
Main	500 iterations	500	31.22ms	0.06ms	baseline
PR	Slow	40	9.25ms	0.23ms	29% fewer calls, 18% faster
PR	Fast	35	5.07ms	0.14ms	22% fewer calls, 12% faster
PR	500 iterations	500	28.35ms	0.06ms	9% faster

Show Preview Performance

Branch	Test Type	Calls	Total Time	Avg per Call	vs Main
Main	Slow	24	2.35ms	0.10ms	baseline
Main	Fast	21	1.76ms	0.08ms	baseline
Main	500 iterations	500	21.49ms	0.04ms	baseline
PR	Slow	24	0.13ms	0.01ms	94% faster
PR	Fast	21	0.09ms	<0.01ms	95% faster
PR	500 iterations	500	2.34ms	<0.01ms	89% faster

[soifou]

Pushed an update, I took another stab at it. Details are in the PR description. LGTM I've been using it for 24 hours and haven't seen any regressions.

[saghen]

Brilliant, thank you!