Problem:
Undo places the cursor wherever the cursor happened to sit at "save
time" (`uh_cursor` is sampled lazily on the first change).
Examples:
- `i` preserves, but `a` does not
- `diw`, `atest<Esc>`, `d^` abandon the original position
- `D`, `o` restore it (by accident).

Solution:
`composite` tracks the pending atom (and its `origin`) across frames.
A `stuffed` continuation frame inherits the `origin` + prepped redo.
Store `origin` info in the undo header, so undo can restore it.

- Not for a mid-command undo break (i_CTRL-G_u).
- Undoing a mapping restores where the mapping started (which
  technically may be different than where the "edit" started).

Problem:
The magic globals `it`, `describe`, etc., are more trouble than they are
worth.

- Hooking into `after_each` requires `getfenv()` hacks.
- They confuse luals/emmylua, because the top-level `.luarc.json` isn't
  merged with `test/.luarc.json` (apparently a luals limitation?)
- They totally defeat discoverability because the user just has to
  "know" about the various magic symbols.

So they harm DX, which means they serve no purpose at all.

Solution:
- Expose the test API from `testutil`, so tests can call `t.it()`,
  `t.describe()`, etc., in the conventional way.
- Drop `getfenv()` hacks.
- Drop the `setfenv()` injection in `load_chunk`.
- Drop `test/_meta.lua`.

Replace the busted-based Lua test runner with a repo-local harness.

The new harness runs spec files directly under `nvim -ll`, ships its own
reporter and lightweight `luassert` shim, and keeps the helper/preload
flow used by the functional and unit test suites.

Keep the file boundary model shallow and busted-like by restoring `_G`,
`package.loaded`, `package.preload`, `arg`, and the process environment
between files, without carrying extra reset APIs or custom assertion
machinery.

Update the build and test entrypoints to use the new runner, add
black-box coverage for the harness itself, and drop the bundled
busted/luacheck dependency path.

AI-assisted: Codex

Problem:
In autocmd examples, using "args" as the event-object name is vague and
may be confused with a user-command.

Solution:
Use "ev" as the conventional event-object name.

Problem: if init_prompt replaces the prompt line at the ': mark, it calls
inserted_bytes with the wrong lnum.

Solution: use the correct lnum. Call appended_lines_mark instead when appending
the prompt at the end.

Problem: prompt_setprompt does not adjust extmarks or trigger on_bytes
buffer-updates when fixing the prompt line.

Solution: adjust them, trigger on_bytes.

Notably, hides extmarks when replacing the entire line (and clearing user
input). Otherwise, when just replacing the prompt text, hides extmarks there,
but moves those after (in the user input area) to the correct spot.

Problem: prompt_setprompt calls coladvance with a byte column, but it expects a
screen column. on_lines buffer-updates aren't fired when fixing the prompt line.

Solution: don't use coladvance. Call changed_lines, which also simplifies the
redraw logic. (and calls changed_cline_bef_curs if needed; added test checks
this)

Unlike https://github.com/neovim/neovim/pull/37743/changes#r2775398744, this
means &modified is set by prompt_setprompt if it fixes the prompt line.
Not setting &modified is inconsistent anyway -- even init_prompt sets it if it
fixes the prompt line.

Problem: `extmark_splice()` was being called before `ml_replace()`,
which caused the on_bytes callback to be invoked with the old buffer
text instead of the new text.

Solution: store metadata for each match in a growing array, call
`ml_replace()` once to update the buffer, then call `extmark_splice()`
once per match.

Closes https://github.com/neovim/neovim/issues/36370.