[wrvsrx]

Accepting a cached UTF-16/UTF-32 completion after typing 三 can turn {三} into {三角洲, deleting the closing brace. Cursor compensation currently adds a byte delta to a range expressed in the client's encoding.

Preserve the request-time line alongside each item's position and compute the cursor delta in the client's encoding before converting the edit to UTF-8. This also handles deletions and IME replacements such as san → 三, where the byte column does not change. Preserve the original line through LSP cache reuse and strip the internal metadata before completionItem/resolve. Byte-based edits retain their existing path, and conversion still happens after compensation to avoid premature clamping on shortened lines.

Closes #2650.

Validation:

• The issue's standalone reproduction fails on 8219b58 and passes with this change.
• 54 deterministic regression cases exercise LSP response processing and edit application across UTF-8/UTF-16/UTF-32, CJK/emoji insertion and deletion, IME replacement, delayed responses, cache reuse, trailing text, and insert/replace/default edit ranges.
• stylua --check . and git diff --check pass.

Run the regression cases with Neovim >= 0.11 and a sibling blink.lib checkout (tested with Neovim 0.12.5+editor-fixes.6 and blink.lib fd9a48e):

nvim --headless -u NONE --cmd 'set rtp+=.,../blink.lib' -l tests/text_edits.lua

The runner stubs only the LSP client; no language server, IME, snippet engine, or native fuzzy build is required. Local checks ran under a 512 MiB cgroup limit with swap disabled.