fix: force UI flush after RPC-created diff on Windows conhost (cmd.exe)

[Cannon07]

Problem

On Windows, when Neovim runs inside the legacy cmd.exe / conhost console, an agent edit would mark the file in neo-tree but the diff tab would not appear until the next edit. Reproduced and root-caused with debug logs: the Lua pipeline (show_diff → mark_change_and_reveal → tab creation) runs identically on every edit — the diff tab is created in Neovim's state each time. The failure is purely a screen flush.

The diff is opened via the hook shim's nvim --headless --server <pipe> --remote-expr … call. The post-tab repaint relied on a single deferred vim.cmd("redraw!"). conhost does not flush that timer-driven redraw when it's triggered from an idle --remote-expr, so the grid is only pushed on the next event (the user's next edit). neo-tree updates meanwhile because it forces its own redraw — hence "marking shows, diff doesn't".

Confirmed terminal-dependent: PowerShell terminal / Windows Terminal → fine; cmd.exe → broken.

Fix

Extract the repaint into a force_redraw() helper. On Windows only, after redraw!:

• nvim__redraw({ flush = true }) — explicitly pushes the grid to the UI, and
• an <Ignore> feedkeys nudge — wakes the input loop so conhost delivers the flush immediately.

Both are pcall-wrapped (nvim__redraw is 0.10+ / semi-private), and the whole block is behind vim.fn.has("win32"), so all other platforms keep byte-identical prior behaviour. Applied to both the inline and side-by-side/tab paths in show_diff.

Not a regression

This is pre-existing Windows behaviour, not introduced by the hook-entry consolidation (#77): the timer_start(10, redraw!) predates Windows support (since #37), and the RPC call path (nvim-call.ps1, --headless) was unchanged by that PR. This fix applies equally to the state already on main.

Testing

• require("code-preview.diff") loads clean under headless nvim.
• diff_lifecycle_spec.lua results are identical before/after this change (verified via stash). Note: that spec currently has pre-existing failures on Windows (config.diff returns nil under minimal_init.lua — a separate Windows test-harness gap, unrelated to this change).
• Manually verified in a cmd terminal: the diff now appears on the first edit.

Suggested manual checks for reviewers (in cmd.exe)

• Single Edit, Write (new file → virtual node), MultiEdit — all paint first try
• Multiple files edited in one turn (several diff tabs)
• Re-edit same file twice (close-then-reopen)
• Accept (post-tool close) and reject (<leader>dq) close cleanly
• PowerShell terminal still works (no regression)

🤖 Generated with Claude Code