[harleypig]

Finishes the editor work left after #127 and #123: the last settings in
OLD_FILES/ move in, undo history is saved, .vimrc works with
XDG_DOTVIM unset, and the :help mappings reference matches the
mappings. One commit per item.

Closes #4
Closes #129
Closes #137
Closes #103
Refs #124
Refs #21
Refs #133
Refs #109

What changed, and why

#4: the rest of OLD_FILES/ moves in, and the directory goes

The mappings move to the late maps.vim and diffopt+=algorithm:patience
moves to set.vim. Each one was checked against what is already mapped
and loaded:

• <Leader>O, <Leader>o, <Leader>y, <Leader>b, <Leader>ml and Y
  were unmapped, and no existing mapping starts with any of them.
• S is Normal mode only. vim-surround's S is a Visual-mode mapping, and
  its Normal-mode keys (ys, yS, cs, ds) are untouched. The built-in
  S does the same as cc, which still works.
• . clashed with vim-repeat. The first time vim-surround calls
  repeat#set, vim-repeat's autoload file maps . to <Plug>(RepeatDot),
  and that silently replaced our mapping. maps.vim now loads that
  autoload file itself and maps . through <Plug>(RepeatDot). Surround
  changes still repeat, and the cursor still goes back to where the change
  started. Without vim-repeat, . falls back to Vim's own repeat. Composing
  the two was chosen over dropping the mapping because it keeps both
  behaviours.
• AppendModeline becomes script-local (s:), since only its mapping
  calls it.

OLD_FILES/ is deleted, along with its mention in .claude/WORKFLOW.md.

#129: undofile

set undofile now sits beside undodir in xdg-settings.vim, so the
existing 90-day clean-up of that directory covers the undo files.

#137: XDG_DOTVIM unset or empty

Decision: fall back, rather than fail loudly. When XDG_DOTVIM is
unset or empty, it now takes the directory .vimrc is in, and a
~/.vimrc symlink is followed to the checkout it points into. .vimrc
always sits at the top of the checkout it configures, so the right
directory is already known, and an error message would only ask the user
to type it in. The variable itself is set, rather than a script-local
copy, so the two set lines stay as they were and anything Vim starts
sees the same value.

#103: the :help mappings reference

• Removed: [g / ]g and <Leader>adr, whose mappings are commented
  out, and the <Tab> / <C-Space> notes for a completion plugin that is
  not installed.
• Added: the mappings moved in from Either merge or remove files in OLD_FILES #4, <Leader>y, <Leader>b, the
  quickfix list (:QuickfixToggle, <Leader>qf, <Leader>q],
  <Leader>q[), and the help window, which opens on the right at no more
  than 80 columns.
• Moved: w!! now has its own "Writing as Root" section saying what it
  expands to. c[ / c] are in the quickfix section, marked as needing
  CoC. The help-buffer <Return> / <C-[> mappings are now listed as this
  config's, not as built-ins. The CoC section now says it needs Node.js.
• The learning-gutter tag, which docs/gutter.md links to, is kept.

Area register (.claude/WORKFLOW.md, Refs #124)

area:filetypes no longer names checkfiletype.vim or the Perl helpers in
tools/, both removed in #124. In area:editor, "diff" (which meant the
removed after/diff.vim) becomes "diff options", meaning the diffopt
this PR adds to set.vim.

coc-html (Refs #21, #133)

coc-html is added to g:coc_global_extensions after coc-git. On the
npm registry it is version 1.9.2, from neoclide/coc-html. It bundles the
HTML language service that the hand-written languageserver.html block in
coc-settings.json tries to provide. That block never starts: it names a
library with no command, and it lists the mustache filetype where
vim-polyglot sets html.mustache. Removing the block from
coc-settings.json lands in the separate intellisense PR, not here.

Three set.vim options that never took effect (Refs #109)

Each was checked with this configuration loaded:

• terse only adds the s flag to shortmess, and the later
  shortmess=a replaces it.
• scroll=11 is reset to half the window height at startup ("Last set
  from changed window size"). In an 8-row terminal it raised E49: Invalid scroll size first.
• wrapmargin=2 is ignored while textwidth=78 is set.

Decision: remove them rather than make them work. Making them work
would bring in behaviour this configuration has never actually had. After
the removal, an 8-row terminal (a real pty, via script) starts without
E49. No automated test was added: the E49 case needs a real
short terminal, and headless vim -es always reports 24 lines.

Testing evaluation

Run to the qa-tester's standard by the PR author; the team-managed
delegation norm does not apply in this repository.

Change	Verdict	Automated	Falsifier: what turns it red
Moved mappings	needs tests: written	tests/vim/maps.bats, 9 new tests driven through :normal	removing any mapping; mapping . with nnoremap again (the vim-surround test goes red); dropping the fallback branch without vim-repeat
diffopt	needs test: written	tests/vim/set.bats	removing the diffopt line
undofile	needs test: written	tests/vim/xdg.bats, which now checks undofile=1 in all 3 XDG cases	removing set undofile
XDG_DOTVIM fallback	needs tests: written	tests/vim/vimrc.bats: set, unset, empty, and a ~/.vimrc symlink	removing the fallback (unset, empty and symlink go red); dropping resolve() (the symlink test goes red, checked by doing it)
learning.txt	no automated test: free-form help text	no; manual note below	none
WORKFLOW.md, coc-html	nothing earns a test: register prose, one list entry	none	none

Every new test was run red against the code before the change, then green
after it. bats tests/vim: 61 ok. bats tests/shell: passes. markdownlint
(CI image): clean.

Entry-point reach: the .vimrc fallback was exercised through the real
entry point (vim -u <path>) in all four states, including a symlinked
~/.vimrc.

Manual verification

1. Where: a shell with XDG_DOTVIM unset, in any directory.
   Do: vim -N -u /path/to/dotvim/.vimrc, then :echo &rtp.
   Success: the first entry is /path/to/dotvim/.vim, and :help learning opens on the right.
   Failure: the first entry is /.vim, or E149 for learning.
2. Do: :help learning-editing.
   Success: "3. Editing" lists <Leader>O, <Leader>o, S, Y, .
   and <Leader>ml.
3. Do: open a file, type foo bar, then 0ysiw", W and ..
   Success: the line reads "foo" "bar".
4. Do: edit a file, :w, quit, reopen it, and press u.
   Success: the edit is undone. The undo file is under
   $XDG_STATE_HOME/vim/undo (or ~/.local/state/vim/undo).

What happened: each step was run headless (vim -es, no terminal), on
this branch:

• Step 1: covered by vimrc.bats; rtp0 and packpath start at the
  checkout's .vim, and :QuickfixToggle exists.
• Step 2: :help jumped to learning-editing, learning-quickfix,
  learning-sudo, learning-help and learning-gutter, each landing on
  its section heading, with no error.
• Step 3: covered by maps.bats; the line reads "aa" "bb", and . is
  still <Plug>(RepeatDot) followed by `[.
• Step 4: after :write, filereadable(undofile(...)) was 1, and the
  file was in state/vim/undo.

The interactive Insert-mode and terminal steps were not run in a real
terminal.

Completion report

1. The functionality works: yes, see the tests and the manual run above.
2. The access model is defined: N/A, a local editor configuration.
3. Permissions are implemented: N/A.
4. Audit logging is implemented: N/A.
5. Validation is complete: the only input is XDG_DOTVIM; unset and empty
   are handled, and a set value is used as given, as before.
6. Errors are handled: the vim-repeat-absent branch falls back to Vim's own
   repeat, and the test suite checks for no errors at startup.
7. Testing is complete: yes, see the testing evaluation.
8. Documentation exists: the :help reference is updated. Not
   complete: README.md still says XDG_DOTVIM must be set, and that
   startup fails without XDG_STATE_HOME/XDG_DATA_HOME, which has been
   false since fix: XDG fallbacks, w!!, paste toggle and ALE linter names (#85, #101, #25, #100) #123. That file is outside this PR. Risk: a reader sets
   variables that are no longer needed; nothing breaks. Next: a README
   pass.
9. The completion report has been produced: this section.

Review

opt-out: no review skill ran. The change refuses no work: it adds no hook,
guard or gate. The judgement calls a reviewer should check are the
composed . mapping, which loads vim-repeat's autoload file at startup
and so also maps u, U and <C-R> from the start rather than on first
use, and the #137 fallback, which sets $XDG_DOTVIM in Vim's environment.

Co-Authored-By: AI Agent noreply@harleypig.com