docs: absorb #234 and #227, discharge the hold, and rule D3 next

The absorption pass after the two merges of 2026-08-11: the LSP
file-watcher fix #234 (ae84d58) and git integration Stage 1 #227
(b867f64), landed in that order so the 2026-08-10 hold on #227 was
discharged rather than overridden.

The handoff anchor moves 9a26ac8 -> b867f64 and lists the ten-merge
first-parent chain between them. Only #227 and #234 are absorbed into
section 1 at this anchor; the eight between keep their facts in their
ledger lanes, several of whose headers still say OPEN --- the anchor
paragraph says so explicitly and repeats the ledger's own rule: trust
the canonical-base line over any lane header.

Both lanes are rewritten to their remainders per rule 4's arc test:

- LSP file watcher: the arc is issue #233 and it stays open until D3
  closes it. The lane now carries the D3 start-state --- the walk
  still recurses into everything every 250ms, no notify dependency in
  the tree, no ignore-list infrastructure to reuse --- and the user's
  ruling that D3 is next.
- Git integration: Stage 2 (gutter markers) is the remainder, and the
  lane preserves the one scheduling constraint that matters: it needs
  new DecorationKind variants, a PROTOCOL_VERSION bump, so it must run
  alone. The five-round review history stays in the PR and framing;
  what the handoff absorbs is what the next lane needs: the
  capture-at-invocation census and its single-mechanism fix, the macOS
  EILSEQ portability fact and its latent crdt-module sibling, the
  root-parsing byte lessons, and the copy/rename presentation ruling.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

COHERENCE.md

@@ -1306,6 +1306,17 @@ Primitive-by-primitive against the list above:
   that found it (§25). All four remain in `lsp.lua`; per §25 the
   symbols are authoritative and the `ad41cf1` line numbers have drifted.
 
+  **Updated again: FIVE, and the fifth is the first outside
+  `lsp.lua`.** Git Stage 1's `*git-status*`
+  (`builtin/runtime/git.lua`) is the concrete evidence P5 asked for
+  that the primitive generalizes past its first consumer — the
+  remediation here was always adoption, not construction. It also
+  added the primitive's one extension: an optional **`keys`** table on
+  the open spec, installed once with the panel's buffer and compared
+  (not re-bound) on reopen, because `Keymap::bind` refuses duplicates
+  and an async consumer re-opens on every refresh. `*buffer-list*` and
+  project-search remain the un-migrated hand-rolled pair.
+
   **`*lsp*` is the only one of the four with a working refresh** — it is
   the only one supplying `on_refresh`. `g` is bound on all four
   unconditionally by `bind_local_keymap`, so the other three carry a
@@ -1435,10 +1446,25 @@ What does not:
 
 - **Code actions apply the first action blindly** — no picker (a
   roadmap "dark matter" item still true at audit).
-- **There is no Git integration at all** — no status, stage, diff,
+- **Git integration reaches status and diff, and no further.** Stage 1
-  blame, or gutter markers anywhere in the tree (gutter git riders and
+  (`docs/git-integration-framing.md`) ships `*git-status*` — a
-  the `ResourceOffer` diff/blame family are named deferrals). The Git
+  `listview` panel over `git status --porcelain=v2 --branch -z`, with
-  affordance list above has nothing to attach to yet.
+  RET visiting the file and `d` showing its file-level diff. There is
+  still **no stage, revert, blame, or gutter marker** anywhere in the
+  tree; gutter git riders need new `DecorationKind` variants (Stage 2,
+  which must be scheduled alone), and the `ResourceOffer` diff/blame
+  family remains a named deferral. The Git affordance list above now has
+  something to attach to; the affordances themselves are unbuilt, and
+  the menu's context vocabulary (`src/menu.rs`) has no `git` context to
+  host them.
+
+  The original audit said "there is no Git integration at all … anywhere
+  in the tree", and that was **literally false when it was written**:
+  `tests/fixtures/pmacs-magit/` is a tracked, installable package that
+  spawns git and parses porcelain v2, with a 32-test acceptance suite
+  (`tests/m8_6_acceptance.rs`). The **product** gap it described was
+  real; the sentence overstated it, and the framing that found the
+  overstatement is the one that closed the gap.
 - No test run/debug affordances (DAP is a future arc,
   `docs/dap-debugging-framing.md`).
 - No missing-tool guidance affordances (§1.2 — the diagnostic that

builtin/runtime/listview.lua

@@ -25,6 +25,7 @@
 --     rows = { { text = "src/foo.rs:12:4", item = <any> }, ... },
 --     on_visit = function(item) ... end,      -- RET/SPC (optional)
 --     on_refresh = function() return rows end, -- g (optional)
+--     keys = { d = "git.diff-file" },          -- extra buffer-local keys
 --   }
 
 pmacs.listview = pmacs.listview or {}
@@ -263,19 +264,193 @@ local function seat_cursor(p, line)
   end
 end
 
+-- The primitive's own key surface, named ONCE so the binder below and
+-- the `keys` validator consult the same list. Previously this was a
+-- sequence of `bind(...)` calls and the set existed nowhere as data,
+-- which is why the git framing had to quote it from the source
+-- (docs/git-integration-framing.md Q#G-7).
+local FIXED_KEYS = {
+  { "RET", "listview.visit" },
+  { "SPC", "listview.visit" },
+  { "n", "cursor.down" },
+  { "<down>", "cursor.down" },
+  { "p", "cursor.up" },
+  { "<up>", "cursor.up" },
+  { "TAB", "listview.toggle" },
+  { "g", "listview.refresh" },
+  { "q", "listview.quit" },
+}
+
 local function bind_local_keymap(buf)
-  local function bind(seq, command)
+  for _, entry in ipairs(FIXED_KEYS) do
-    pmacs.keymap.bind { scope = "buffer", buffer = buf, sequence = seq, command = command }
+    pmacs.keymap.bind {
+      scope = "buffer", buffer = buf, sequence = entry[1], command = entry[2],
+    }
   end
-  bind("RET", "listview.visit")
+end
-  bind("SPC", "listview.visit")
+
-  bind("n", "cursor.down")
+-- ---------------------------------------------------------------------
-  bind("<down>", "cursor.down")
+-- Consumer-supplied keys (Q#G-7)
-  bind("p", "cursor.up")
+-- ---------------------------------------------------------------------
-  bind("<up>", "cursor.up")
+--
-  bind("TAB", "listview.toggle")
+-- An optional `keys = { <sequence> = <command name> }` on the open
-  bind("g", "listview.refresh")
+-- spec, bound through the SAME `pmacs.keymap.bind { scope = "buffer" }`
-  bind("q", "listview.quit")
+-- path as the fixed set above. It exists because a consumer cannot
+-- safely bind its own key from outside: `open` disambiguates a name
+-- collision to `<2>`, so the name a consumer passed is not necessarily
+-- the buffer it got, and this module is the only place the handle is
+-- known. No key is intercepted anywhere — COHERENCE.md §6's shadow
+-- count is unchanged by this.
+--
+-- INSTALL-ONCE, MATCH-ON-REOPEN. `Keymap::bind` refuses duplicates
+-- (`KeymapError::DuplicateBinding`, "Refuse rather than silently
+-- overwrite", src/keymap_tree.rs), and a consumer built on the async
+-- completion model calls `open` again on EVERY refresh. So keys are
+-- installed when the buffer is created and a later `open` for a live
+-- panel does not re-bind — it COMPARES, and errors on divergence.
+-- Silently keeping the old binding would hand the consumer a key that
+-- does something other than what it just asked for, which is the dead-
+-- or-lying-key defect this module already condemns for `g`.
+
+-- A key sequence's whitespace-separated chord tokens. That is exactly
+-- how `parse_sequence` (src/key.rs) splits one, so a prefix relation
+-- computed here is the same relation the trie would find.
+local function chords_of(sequence)
+  local out = {}
+  for token in sequence:gmatch("%S+") do out[#out + 1] = token end
+  return out
+end
+
+-- True when one chord list is a STRICT prefix of the other. Either
+-- direction is a conflict: `Keymap` refuses both turning a leaf into a
+-- submap (`WouldExtendLeaf`) and shadowing a submap with a leaf
+-- (`WouldShadowSubmap`), and a `keys` table must not be able to reach
+-- either.
+local function prefix_conflict(a, b)
+  local short, long = a, b
+  if #a > #b then short, long = b, a end
+  if #short == 0 or #short == #long then return false end
+  for i = 1, #short do
+    if short[i] ~= long[i] then return false end
+  end
+  return true
+end
+
+-- Normalize `keys` into a sorted array of `{ sequence, command }`.
+-- Sorted so the comparison on reopen and every error message are
+-- deterministic (`pairs` order is not).
+local function normalized_keys(keys)
+  if keys == nil then return {} end
+  if type(keys) ~= "table" then
+    error(string.format(
+      "listview: `keys` must be a table of sequence -> command name; got %s",
+      type(keys)))
+  end
+  local out = {}
+  for sequence, command in pairs(keys) do
+    if type(sequence) ~= "string" or sequence == "" then
+      error("listview: every `keys` entry must be keyed by a non-empty key sequence")
+    end
+    if type(command) ~= "string" or command == "" then
+      error(string.format(
+        "listview: `keys[%q]` must be a command NAME (a non-empty string); got %s",
+        sequence, type(command)))
+    end
+    out[#out + 1] = { sequence = sequence, command = command }
+  end
+  table.sort(out, function(a, b) return a.sequence < b.sequence end)
+  return out
+end
+
+-- A FIRST-PASS collision check, for a better message than the keymap's.
+--
+-- It compares RAW TOKENS, and that is deliberately not sufficient: the
+-- key parser canonicalizes aliases before it ever reaches the trie
+-- (`parse_key_code`, src/key.rs, uppercases and folds `RET`/`RETURN`/
+-- `ENTER`, `SPC`/`SPACE`, `ESC`/`ESCAPE`, `BS`/`BACKSPACE`,
+-- `DEL`/`DELETE`), so `keys = { RETURN = ... }` is a collision this
+-- function cannot see.
+--
+-- **`Keymap::bind` is the authority, and `ensure_panel` tears the panel
+-- down when it refuses.** That is not a fallback for a check that
+-- happens to be weak --- it is the only version that cannot go stale. A
+-- Lua-side canonicalizer would be a second copy of `parse_key_code`'s
+-- alias table, and the day the Rust one gains a name the Lua one would
+-- silently stop seeing that alias, reintroducing exactly this bug for
+-- it. (There is also no way to canonicalize an arbitrary sequence from
+-- Lua today: `display_sequence` is reachable only through
+-- `describe.key` and `keymap.list`, which both require the sequence to
+-- be BOUND already.)
+--
+-- So what this buys is diagnosis, not safety: a named "that is the
+-- panel's own `g`" instead of a raw `DuplicateBinding`.
+local function check_key_collisions(entries)
+  for i, entry in ipairs(entries) do
+    local mine = chords_of(entry.sequence)
+    for _, fixed in ipairs(FIXED_KEYS) do
+      if entry.sequence == fixed[1] then
+        error(string.format(
+          "listview: `keys` may not rebind %q --- it is part of the panel's "
+          .. "own key surface (RET SPC n <down> p <up> TAB g q), bound to %q",
+          entry.sequence, fixed[2]))
+      end
+      if prefix_conflict(mine, chords_of(fixed[1])) then
+        error(string.format(
+          "listview: `keys` entry %q conflicts with the panel's own %q --- "
+          .. "one is a prefix of the other, which the keymap refuses rather "
+          .. "than turning a binding into a submap",
+          entry.sequence, fixed[1]))
+      end
+    end
+    for j = i + 1, #entries do
+      if prefix_conflict(mine, chords_of(entries[j].sequence)) then
+        error(string.format(
+          "listview: `keys` entries %q and %q conflict --- one is a prefix "
+          .. "of the other", entry.sequence, entries[j].sequence))
+      end
+    end
+  end
+end
+
+-- Bind the entries, naming which one the keymap refused.
+--
+-- It does NOT roll back the keys it already bound: its caller owns
+-- teardown, and the caller's teardown is killing the whole buffer,
+-- which takes the buffer's entire keymap scope with it
+-- (`after_buffer_removed` -> `KeymapStack::remove_buffer`). Unbinding
+-- here as well would be a second, weaker cleanup mechanism for the same
+-- failure --- and the weaker one is what let a half-built panel survive.
+local function install_keys(buf, entries)
+  for _, entry in ipairs(entries) do
+    local ok, err = pcall(pmacs.keymap.bind, {
+      scope = "buffer", buffer = buf,
+      sequence = entry.sequence, command = entry.command,
+    })
+    if not ok then
+      error(string.format(
+        "listview: cannot bind %q to %q: %s",
+        entry.sequence, entry.command, tostring(err)))
+    end
+  end
+end
+
+local function keys_match(a, b)
+  if #a ~= #b then return false end
+  for i = 1, #a do
+    if a[i].sequence ~= b[i].sequence or a[i].command ~= b[i].command then
+      return false
+    end
+  end
+  return true
+end
+
+local function render_keys(entries)
+  if #entries == 0 then return "none" end
+  local parts = {}
+  for i, entry in ipairs(entries) do
+    parts[i] = string.format("%s=%s", entry.sequence, entry.command)
+  end
+  return table.concat(parts, " ")
 end
 
 -- Build the persistent panel record for `name`. A user-killed panel
@@ -293,9 +468,23 @@ end
 -- collision disambiguates `<2>`..`<99>`, and exhausting the limit raises
 -- rather than adopting --- the rule terminal.lua:300-305 states and
 -- dired.lua:476-504 already implements.
-local function ensure_panel(name)
+local function ensure_panel(name, key_entries)
   local p = panel_for_requested_name(name)
-  if p then return p end
+  if p then
+    -- Match-on-reopen (Q#G-7). A live panel keeps the keys it was
+    -- created with; a DIFFERENT table is a consumer asking for
+    -- something it will not get, so it is an error rather than a
+    -- silently ignored request.
+    if not keys_match(p.keys, key_entries) then
+      error(string.format(
+        "listview: %s is already open with keys [%s]; this open asks for "
+        .. "[%s]. Keys are installed once with the panel's buffer, so the "
+        .. "second table would be silently ignored --- close the panel "
+        .. "first, or pass the same keys",
+        name, render_keys(p.keys), render_keys(key_entries)))
+    end
+    return p
+  end
 
   local actual = name
   if find_buffer_by_name(actual) then
@@ -315,29 +504,64 @@ local function ensure_panel(name)
 
   local buf = pmacs.buffer.create(actual)
   p = { requested_name = name, buffer = buf, line_to_item = {},
-        line_to_row = {}, collapsed = {}, rows = {}, visible = 0 }
+        line_to_row = {}, collapsed = {}, rows = {}, visible = 0,
-  panels[#panels + 1] = p
+        keys = key_entries }
-  -- Read-only (Q#P3): every non-bypass edit is rejected, with a NAMED
+  -- ALL-OR-NOTHING from here. Everything below mutates a buffer that
-  -- error. Kept beside the rope lock, not replaced by it: the layering
+  -- does not yet belong to a panel, and `install_keys` can genuinely
-  -- at terminal.lua:351-366 --- the rope lock protects the daemon copy,
+  -- fail: the raw-token preflight cannot see an alias spelling of a
-  -- this and the round-trip mark protect a semantic frontend's own
+  -- fixed key (`RETURN` for `RET`), so `Keymap::bind` is the first thing
-  -- mirror, and neither substitutes for the other. The intercept lives
+  -- to notice, and by then the buffer exists, carries a read-only
-  -- as long as the buffer; no teardown (the buffer-list precedent for
+  -- intercept and a round-trip mark, and holds the fixed keymap.
-  -- its keymap).
+  --
-  pmacs.buffer.add_intercept(buf, function()
+  -- Leaving it behind is worse than it sounds: it is read-only, it is in
-    error(actual .. " is read-only")
+  -- no `panels` record so nothing owns or can reach it, and the next
+  -- `open` for the same name finds it by name and disambiguates itself
+  -- to `<2>` --- so a rejected `keys` table silently renames the panel.
+  local built, err = pcall(function()
+    -- Read-only (Q#P3): every non-bypass edit is rejected, with a NAMED
+    -- error. Kept beside the rope lock, not replaced by it: the layering
+    -- at terminal.lua:351-366 --- the rope lock protects the daemon copy,
+    -- this and the round-trip mark protect a semantic frontend's own
+    -- mirror, and neither substitutes for the other. The intercept lives
+    -- as long as the buffer; no teardown (the buffer-list precedent for
+    -- its keymap).
+    pmacs.buffer.add_intercept(buf, function()
+      error(actual .. " is read-only")
+    end)
+    -- Q#P6: semantic frontends must round-trip keys while this panel
+    -- is focused (RET = visit, not an optimistic newline).
+    pmacs.buffer.set_round_trip_input(buf, true)
+    bind_local_keymap(buf)
+    install_keys(buf, key_entries)
   end)
-  -- Q#P6: semantic frontends must round-trip keys while this panel
+  if not built then
-  -- is focused (RET = visit, not an optimistic newline).
+    -- `kill` is the whole teardown, not a convenience: it removes the
-  pmacs.buffer.set_round_trip_input(buf, true)
+    -- buffer AND, through `after_buffer_removed`, prunes the buffer's
-  bind_local_keymap(buf)
+    -- keymap scope, its config locals and its folds. Unbinding key by
+    -- key would leave the buffer itself --- which is the defect.
+    pcall(pmacs.buffer.kill, buf)
+    -- Level 0: re-raise the inner message verbatim rather than stacking
+    -- this line's position onto it.
+    error(err, 0)
+  end
+  -- Registered LAST, deliberately: a failure above must leave no record
+  -- claiming keys it did not bind. Nothing above needs the panel to be
+  -- in `panels` --- the intercept, the round-trip mark and the keymap
+  -- all address the buffer directly.
+  panels[#panels + 1] = p
   return p
 end
 
 function pmacs.listview.open(spec)
   assert(type(spec) == "table" and type(spec.name) == "string",
     "listview.open: spec.name (string) required")
-  local p = ensure_panel(spec.name)
+  -- The cheap checks first, so the common mistakes are named before
+  -- anything is created. The ones this pass cannot see --- alias
+  -- spellings --- are caught by `Keymap::bind` inside `ensure_panel`,
+  -- which tears the panel down rather than leaving it half-built.
+  local key_entries = normalized_keys(spec.keys)
+  check_key_collisions(key_entries)
+  local p = ensure_panel(spec.name, key_entries)
   p.header = spec.header or spec.name
   p.on_visit = spec.on_visit
   p.on_refresh = spec.on_refresh

builtin/runtime/lsp.lua

@@ -1924,7 +1924,8 @@ end
 local FILE_WATCH_INTERVAL_MS = 250
 
 -- file_watchers[tostring(sid)][registrationId] = list of watch records
--- ({ cancelled = bool, _sleep = handle? }), one per glob watcher.
+-- ({ cancelled = bool, form = "relative"|"absolute", _sleep = handle? }),
+-- one per glob watcher.
 local file_watchers = {}
 
 -- WatchKind is a bitmask (Create=1, Change=2, Delete=4); test it
@@ -2058,7 +2059,18 @@ end
 local FC_CREATED, FC_CHANGED, FC_DELETED = 1, 2, 3
 
 local function start_file_watcher(sid, base, glob, kind_mask, record)
-  local matches = glob_matcher(glob)
+  -- Per LSP, a plain-string glob matches the file's ABSOLUTE path,
+  -- while a RelativePattern's pattern is relative to its base — the
+  -- record's `form` (from resolve_watcher) picks the match subject.
+  -- scan_tree always walks in relative terms; only the string handed
+  -- to the matcher changes.
+  local match_glob = glob_matcher(glob)
+  local matches = match_glob
+  if record.form == "absolute" then
+    matches = function(rel)
+      return match_glob(base .. "/" .. rel)
+    end
+  end
   pmacs.async(function()
     local prev = scan_tree(base, matches)
     while not record.cancelled and server_is_live(sid) do
@@ -2069,6 +2081,29 @@ local function start_file_watcher(sid, base, glob, kind_mask, record)
       if record.cancelled or not server_is_live(sid) then break end
 
       local cur = scan_tree(base, matches)
+      -- The seam that makes the recheck below WITNESSABLE. `scan_tree`
+      -- suspends on `read_dir` once per directory, and the race is a
+      -- cancel arriving during one of those suspensions --- which no
+      -- arrangement of real timing can be made to happen on demand.
+      -- Same reason `git.lua` exposes `_deliver_status`: the contract is
+      -- about an interleaving the caller does not choose. Unset in
+      -- production, so this costs one nil test per tick.
+      -- `cur` is handed over so a test can cancel on THE SCAN THAT
+      -- OBSERVED a given change. Cancelling on any other scan is not a
+      -- witness: the loop would break at the post-sleep check on the
+      -- next iteration and emit nothing anyway, so the assertion would
+      -- pass with the recheck below deleted.
+      if pmacs.lsp._after_scan_for_tests then
+        pcall(pmacs.lsp._after_scan_for_tests, record, cur)
+      end
+      -- RECHECKED AFTER THE SCAN, not only after the sleep (review P2).
+      -- The coroutine is suspended for most of a tick with `_sleep`
+      -- already cleared, so a cancel landing there sets `cancelled` and
+      -- has no sleep to interrupt. Without this line the resumed scan
+      -- runs on to `did_change_watched_files` below and a SUPERSEDED
+      -- watcher emits one last batch under its OLD pattern. One batch is
+      -- enough: it is a wrong-pattern notification the server acts on.
+      if record.cancelled or not server_is_live(sid) then break end
       local changes = {}
       for rel, sig in pairs(cur) do
         local was = prev[rel]
@@ -2097,22 +2132,44 @@ local function start_file_watcher(sid, base, glob, kind_mask, record)
 end
 
 -- Resolve a GlobPattern (string | { baseUri, pattern }) to
--- (base_dir, pattern). A bare string with no base falls back to the
+-- (base_dir, pattern, form). The form must travel with the pair: a
--- directory of an attached file on `sid` (best effort).
+-- RelativePattern's pattern is relative to its baseUri, and dropping
+-- that distinction is what made absolute server globs unable to match
+-- anything. A bare string with no base falls back to the directory of
+-- an attached file on `sid` (best effort).
+--
+-- THE FORM COMES FROM THE PATTERN, NOT FROM THE UNION ARM (review P1).
+-- The first fix for #233 returned `"absolute"` for every string, which
+-- is a different bug wearing the same shape: LSP 3.17 defines `Pattern`
+-- relative to a base path, and VS Code treats a string watcher as
+-- applying across workspace folders, so a bare `*.txt` is a VALID
+-- relative pattern. Classifying it absolute matched it against
+-- `<base>/foo.txt`, which `^[^/]*%.txt$` can never match --- so that
+-- fix silently broke a case that worked before it. A leading `/` is
+-- what makes a pattern absolute; the arm it arrived in is not.
 local function resolve_watcher(sid, gp)
   if type(gp) == "table" and gp.baseUri then
-    return pmacs.lsp.path_for_uri(gp.baseUri), gp.pattern or "**"
+    return pmacs.lsp.path_for_uri(gp.baseUri), gp.pattern or "**", "relative"
   end
   if type(gp) == "string" then
     for _, rec in pairs(attachments) do
       if rec.server == sid and rec.uri then
         local p = pmacs.lsp.path_for_uri(rec.uri)
         local dir = p and p:match("^(.*)/[^/]*$")
-        if dir then return dir, gp end
+        if dir then
+          return dir, gp, (gp:sub(1, 1) == "/") and "absolute" or "relative"
+        end
       end
     end
   end
-  return nil, nil
+  return nil, nil, nil
+end
+
+local function cancel_watch_records(recs)
+  for _, r in ipairs(recs or {}) do
+    r.cancelled = true
+    if r._sleep then pcall(function() r._sleep:cancel() end) end
+  end
 end
 
 local function register_file_watchers(sid, registrations)
@@ -2120,11 +2177,16 @@ local function register_file_watchers(sid, registrations)
   file_watchers[skey] = file_watchers[skey] or {}
   for _, reg in ipairs(registrations or {}) do
     if reg.method == "workspace/didChangeWatchedFiles" then
+      -- Re-registering a live id supersedes it (rust-analyzer does
+      -- this): cancel the outgoing records first, because the table
+      -- write below drops the only reference to them and an
+      -- uncancelled record polls until the server dies.
+      cancel_watch_records(file_watchers[skey][reg.id])
       local recs = {}
       for _, w in ipairs((reg.registerOptions or {}).watchers or {}) do
-        local base, pat = resolve_watcher(sid, w.globPattern)
+        local base, pat, form = resolve_watcher(sid, w.globPattern)
         if base and pat then
-          local r = { cancelled = false }
+          local r = { cancelled = false, form = form }
           recs[#recs + 1] = r
           start_file_watcher(sid, base, pat, w.kind or 7, r)
         end
@@ -2139,10 +2201,7 @@ local function unregister_file_watchers(sid, unregs)
   if not byid then return end
   for _, u in ipairs(unregs or {}) do
     if u.method == "workspace/didChangeWatchedFiles" and byid[u.id] then
-      for _, r in ipairs(byid[u.id]) do
+      cancel_watch_records(byid[u.id])
-        r.cancelled = true
-        if r._sleep then pcall(function() r._sleep:cancel() end) end
-      end
       byid[u.id] = nil
     end
   end

docs/active-work.md

@@ -5,6 +5,21 @@ landed on `main`. Read it after `docs/agent-handoff.md`. Remove completed
 entries when their PR merges; do not let this become a second permanent
 backlog.
 
+**Updated 2026-08-11 — two merges and a discharge.** The LSP file
+watcher D1+D2 landed as **#234** (`ae84d58`) after one review round
+(P1 form-from-the-pattern, P2 cancelled-scan emission — both
+bite-verified), and git integration Stage 1 landed as **#227**
+(`b867f64`) immediately after, refreshed onto the merged base and
+re-gated. The 2026-08-10 hold on #227 is **discharged in order**, not
+overridden. Both lanes below are rewritten to their remainders (D3;
+git Stage 2), their durable facts absorbed into
+`docs/agent-handoff.md` §1. **Next, by user ruling: D3.** The
+canonical-base line below and the handoff anchor both moved to
+`b867f64`. Several other lane headers still say OPEN for PRs that have
+since merged (#224–#232) — per this file's own rule, trust the
+canonical-base line over any lane header; those absorptions remain
+owed by their own lanes.
+
 **Updated later the same day, on a new machine.** Development moved to
 the laptop; the recovery path in "Repository authority" below was
 exercised from this checkout and the `githubsucks` alias was absent and
@@ -111,7 +126,14 @@ lesson, §1 for the two framings).
   are identical on every machine. Remote names are otherwise
   machine-local: `origin` may name this canonical URL, a release mirror,
   or something else, and therefore has no authority by name alone.
-- Canonical base at this snapshot: **`githubsucks/main` @ `9a26ac8`** —
+- Canonical base at this snapshot: **`githubsucks/main` @ `b867f64`** —
+  git integration Stage 1 **#227**, atop `ae84d58` the LSP file-watcher
+  fix **#234**, atop `0e4c58d` destination capture **#231**, `3cc1b85`
+  worker identity Stage 1 **#232**, `0857bf4` discovery Stage 2
+  **#228**, `0190102` LSP LaTeX coverage **#230**, `7cf4653` the gate
+  `--protocol` build step **#229**, `4bc55e8` per-worktree gate target
+  dirs **#225**, `dcb852e` the R8 fixture fix **#226** and `b833b13`
+  the QoL docs retirement **#224**. Beneath those, `9a26ac8`:
   GPU horizontal scroll **#223**, which **closes the QoL arc**, atop
   `2b56d16` TUI horizontal scroll **#222**, `02f3ec3` `ui.line-wrap`
   **#221** (protocol v22), `218d2e7` GUI zoom **#220** and `da56bec`
@@ -210,6 +232,43 @@ hazard in a shape that looks committed. **A documented error message
 that never appears is worse than no documentation**, because the reader
 waits for a signal that is not coming.
 
+## LSP file watcher (issue #233) — D1+D2 MERGED as #234; D3 IS NEXT
+
+**Issue #233** — https://github.com/levineuwirth/pmacs/issues/233,
+still OPEN: it closes when D3 does. **PR #234 MERGED 2026-08-11**
+(`main` @ `ae84d58`), one review round. The framing is
+`docs/lsp-file-watcher-framing.md`, revision 2 — it carries the full
+record: the approved design, the answered ruling, and the two review
+findings (P1 form-from-the-pattern, P2 cancelled-scan emission) with
+their bite results. Durable facts are absorbed in
+`docs/agent-handoff.md` §1.
+
+**The #227 hold is DISCHARGED.** The 2026-08-10 ruling held #227
+unmerged until this was resolved; #234 merged first and #227 followed
+the same day (`b867f64`), refreshed and re-gated on the merged base.
+
+**D3 — the polling cost — is the remainder, and the user has ruled it
+is next (2026-08-11).** No branch and no framing yet. What is known,
+verified while framing D1/D2:
+
+- After #234 the watcher is *correct* but still walks: `walk` recurses
+  unconditionally and `matches` gates only recording, so rust-analyzer
+  walks the whole tree — `.git`, `target`, `node_modules` included —
+  every 250 ms, six times per tick (was twelve before D2), one async
+  job per directory. The modeline still shows the churn, at roughly
+  half the pre-#234 rate.
+- **No `notify`/inotify dependency in the tree** — a real
+  filesystem-notification primitive is a new crate plus a new Rust
+  primitive plus its Lua binding.
+- **No ignore-list infrastructure to reuse** — `src/project.rs` knows
+  `.git` as a *marker* name, not as something to skip.
+- Options named in the issue: coalesce a server's watchers into one
+  scan; root the scan at the workspace root; an ignore list; back off
+  when nothing changes; a real notification primitive.
+- It is a `COHERENCE.md` §9 concern — background work with no
+  ownership model — and the activity indicator that surfaced it is
+  §9's own Stage 1. The framing must state its §20 coherence impact.
+
 ## `scripts/gate` — PR #225 OPEN (build tooling)
 
 **PR #225** — https://github.com/levineuwirth/pmacs/pull/225. Written
@@ -265,6 +324,40 @@ also removed: this branch's "R8 NEEDS A LANE" investigation block, and
 durable facts are in the retired registry row and the handoff §6
 census.
 
+## Git integration — STAGE 1 MERGED as #227; Stage 2 must be scheduled alone
+
+**PR #227 MERGED 2026-08-11** (`main` @ `b867f64`), after five review
+rounds, a macOS CI round, and a base refresh: it was held unmerged
+behind issue #233 by the 2026-08-10 ruling, refreshed onto the merged
+base (`e2394c7`, a clean merge whose only file shared with #234 was
+this ledger), re-gated 11/11 locally and 14/14 on CI. Durable facts
+are absorbed in `docs/agent-handoff.md` §1; the framing
+(`docs/git-integration-framing.md`, revision 5) and the PR carry the
+full five-round review history.
+
+**What shipped:** `*git-status*` — a `listview` panel over
+`git --no-optional-locks -C <dir> status --porcelain=v2 --branch -z` —
+and `*git-diff*` (file-level, plain generated text), with the
+install-once `keys` extension on `listview`. `builtin/runtime/git.lua`,
+34 acceptance tests, **no wire change**.
+
+**Stage 2 (gutter markers) is the remainder, and it is NOT freely
+schedulable: it needs new `DecorationKind` variants — a
+`PROTOCOL_VERSION` bump — so it must run alone**, per the strict
+serialization rule on wire changes. No branch, no framing yet.
+
+**Residue that stays live here:**
+
+- **§9 negative impact stands:** git runs as a spawned process, and
+  spawned processes do not appear in `*workers*`. A fifth
+  unattributable background thing, labelled honestly; a label is not
+  attribution. The D3 lane (above) and §9 Stage 2 own the model.
+- **The latent macOS sibling:** `tests/gpu_invocation_acceptance.rs`
+  writes a non-UTF-8 filename to disk inside
+  `#[cfg(feature = "crdt")]`, and the crdt job is ubuntu-only — it
+  fails the day that job gains a macOS leg, the same way `g6_2` did
+  (handoff §1: macOS cannot hold a non-UTF-8 filename).
+
 ## Destination capture (Q#JR14 generalization) — PR #231 OPEN, revision 9, cleared to merge
 
 **PR #231** — https://github.com/levineuwirth/pmacs/pull/231. #227
@@ -1397,6 +1490,7 @@ authoritative tip** — the ref, not a SHA. Recover with
   — added in the second round — a **rename of either** the build or the
   sweep step each fail the suite.
 
+
 ## QoL arc retirement — PR #224 OPEN (docs only)
 
 **PR #224** — https://github.com/levineuwirth/pmacs/pull/224. Written

docs/agent-handoff.md

@@ -1,6 +1,20 @@
 # Agent handoff — cross-machine continuity
 
-**Last updated: 2026-08-08.** `main` is **`9a26ac8`** — GPU horizontal
+**Last updated: 2026-08-11.** `main` is **`b867f64`** — git integration
+Stage 1 **#227** (`*git-status*` / `*git-diff*`, no wire change), atop
+`ae84d58` **#234**, the LSP file-watcher correctness fix (issue #233
+D1+D2, one review round; **D3 — the polling cost — is deliberately
+unfixed and is the ruled next lane**). Beneath them, in first-parent
+order: **#231** destination capture, **#232** worker identity Stage 1,
+**#228** discovery Stage 2, **#230** LSP LaTeX coverage, **#229** the
+gate `--protocol` build step, **#225** per-worktree gate target dirs,
+**#226** the R8 fixture fix, and **#224** the QoL docs retirement.
+**Only #227 and #234 are absorbed into §1 at this anchor**; the eight
+between carry their facts in their `docs/active-work.md` lanes, several
+of whose headers still say OPEN — trust this chain over any lane
+header, per the ledger's own rule.
+
+Previously **2026-08-08**: `main` was `9a26ac8` — GPU horizontal
 scroll **#223**, which **closes the QoL arc** (§1). Beneath it the arc's
 other four: **#222** TUI horizontal scroll, **#221** `ui.line-wrap` at
 protocol v22, **#220** GUI zoom, **#219** `full_grid` honored by the
@@ -86,8 +100,94 @@ reads it the way you just did.
 For volatile branches, checkpoints, verification, and recovery
 commands, read `docs/active-work.md` immediately after this file.
 
-## 1. Where the project stands (2026-08-08)
+## 1. Where the project stands (2026-08-11)
 
+- **Git integration Stage 1 — MERGED as #227 (2026-08-11).**
+  `*git-status*` (a `listview` panel over `git --no-optional-locks -C
+  <dir> status --porcelain=v2 --branch -z`) and `*git-diff*`
+  (file-level, plain generated text), plus an install-once `keys`
+  extension on `listview`. No wire change; **Stage 2 (gutter markers)
+  needs new `DecorationKind` variants — a `PROTOCOL_VERSION` bump — and
+  must be scheduled alone.** Held unmerged behind issue #233 by user
+  ruling, then landed the same day as #234, refreshed and re-gated on
+  the merged base. Durable facts:
+  - **Capture at invocation, never read at continuation — enforced by
+    ONE mechanism, not four counters.** Review found FOUR instances of
+    the same shape (ordering by `rev-parse` completion; `state.root`
+    read mid-plan; no generation on the diff path; one byte inside a
+    fix). `new_channel()` tickets answer "is this still the request in
+    force?" — **two channels, deliberately**: status and diff are
+    independent things a user asks for, so each gets its own ordering
+    and what is shared is the mechanism. The async-continuation census
+    (three continuations, one dispatcher, one synchronous impostor)
+    lives in the PR and framing.
+  - **macOS cannot hold a non-UTF-8 filename, and this project will
+    hit it again.** APFS/HFS+ reject invalid UTF-8 at the syscall
+    (`EILSEQ`, errno 92); Linux's VFS treats names as opaque bytes. It
+    cannot be reached around the filesystem either: an index-only
+    entry still fails `git status`'s lstat with `EILSEQ`, which git
+    skips — **there is no macOS arrangement in which real `git status`
+    names a non-UTF-8 path.** The fix pattern: split coverage along
+    the line the platform draws — behaviour runs everywhere (rows
+    supplied through the `_deliver_status` seam), only *provenance* is
+    Linux-gated. A latent sibling in `gpu_invocation_acceptance`'s
+    crdt module is recorded in the git lane.
+  - **Both root-parsing bugs lived in a pattern.** `rev-parse
+    --show-toplevel` gets exactly ONE trailing `\n` removed, by an
+    explicit last-byte test: `\r` is a legal POSIX name byte, and
+    `rev-parse` has **no `-z`** — it echoes a literal `-z` onto stdout
+    with exit 0, checked against the installed git rather than
+    assumed. `first_line` was deliberately left alone both times: its
+    three callers feed the single-line status band, where truncation
+    is right.
+  - **`{:?}` on a string containing NUL cannot build a `-z` fixture**
+    — Lua's decimal escape swallows following digits, so payloads are
+    assembled as raw bytes with three-digit escapes, joined in Lua
+    with `string.char(0)`.
+  - **A copy renders `copied from`, but `kind` stays `"rename"` for
+    both, deliberately.** Every behaviour keyed on it is identical;
+    splitting would force every present and future consumer to spell
+    both arms, and a forgotten arm silently degrades copies. The score
+    byte carries the distinction where presentation needs it.
+- **LSP file watcher — D1+D2 MERGED as #234 (2026-08-11); D3 is
+  next.** Issue #233: with any server that dynamically registers
+  `workspace/didChangeWatchedFiles`, plain-string globs never matched
+  (matched **relative** where LSP says **absolute** — rust-analyzer
+  saw no file change, ever; gopls saw `go.mod` but never `.go`), and
+  re-registering a live id leaked the previous pollers uncancellably
+  (rust-analyzer registers twice under one id: 12 pollers, 6
+  unreachable). Invisible for three months until #232's activity
+  indicator — §9's instrument doing exactly its job. Durable facts:
+  - **The GlobPattern form travels with the pattern, and it is read
+    FROM the pattern, not from the union arm.** `resolve_watcher`
+    returns `(base, pattern, form)`; a leading `/` is what makes a
+    string absolute. The first fix classified every string absolute —
+    repairing rust-analyzer while silently breaking bare `*.txt`, a
+    case that had worked since May. Review caught it (P1).
+  - **A scan that completes after cancellation must not emit (P2).**
+    The watcher coroutine spends most of a tick suspended in
+    `read_dir` awaits with `_sleep` already cleared, so a cancel
+    landing there had nothing to interrupt and the resumed scan
+    emitted one stale batch under the superseded pattern. Cancellation
+    and liveness are rechecked after the scan;
+    `pmacs.lsp._after_scan_for_tests` (nil in production, handed the
+    scan result) exists because no real timing produces that
+    interleaving on demand — `git.lua`'s `_deliver_status` device
+    again.
+  - **F1's lesson fired twice in one lane.** The pre-existing test was
+    insensitive (`**/` compiles to `.-`, which spans `/`, so it passes
+    under either match subject) — and then the lane's own flat-pattern
+    guard constrained the RelativePattern *object* arm while P1's
+    regression lived in the *string* arm. A guard proves things about
+    the arm it exercises, nothing more.
+  - **All six watcher tests are mutation-verified, each bite failing
+    only its own defect** — the two review fixes re-verified
+    independently after review.
+  - **D3 is deliberately unfixed and ruled next**: the walk still
+    recurses into everything every 250 ms, six jobs per tick for
+    rust-analyzer. The D3 lane in `docs/active-work.md` carries what
+    was checked (no notify dependency, no ignore-list infrastructure)
+    and the option space.
 - **QoL arc — CLOSED. All five stages merged (#219, #220, #221, #222,
   #223).** From one daily-driver report: terminal zoom broke TUI
   rendering and did nothing in the GUI, and a long line was unreadable

docs/git-integration-framing.md

@@ -0,0 +1,708 @@
+# Git integration — Stage 1: seeing what changed
+
+**Status: revision 5, APPROVED 2026-08-09. Implementation may
+proceed.**
+
+**Revision 5 completes the unborn-repository policy, which revision 4
+wrote as three disjoint rows when a single file can be in two states at
+once.** `AM` — staged, then edited again — is not exotic; it is what a
+first commit looks like halfway through. The states below were
+**enumerated from a real unborn repository**, not reasoned about, and
+one of them settles a case by ruling it out entirely.
+
+**Revision 4 fixes two contracts that would have failed in ordinary
+use, both verified against real behaviour rather than reasoned about:**
+re-binding `d` on every refresh (keymap binds refuse duplicates, so
+*every successful refresh* would have errored), and two git exit states
+the failure predicate got wrong. Measured, not assumed — the exit codes
+below were produced in a scratch repository.
+
+**Revision 3 pins four Stage 1 contracts revision 2 left loose, and two
+of those were again claims I made without reading the code I was
+crediting.** I attributed selection preservation to `listview` and
+`d` to its key surface; neither is true, and both were checkable in the
+file I had already cited. The pattern is worth naming since it has now
+recurred across three revisions: **I cite a file, then describe what I
+expect it to contain.**
+
+**Revision 2 answers four blockers, two of which were factual errors in
+revision 1 that scouting should have caught and did not.** I read
+`ProjectKind::Git`'s name instead of its doc comment, and I quoted
+`COHERENCE.md` §15's "no Git integration anywhere in the tree" without
+checking whether it was still true of the tree. It is not.
+
+---
+
+## 1. Why this, and why now
+
+`COHERENCE.md` §15 is blunt about it:
+
+> **There is no Git integration at all** — no status, stage, diff,
+> blame, or gutter markers anywhere in the tree (gutter git riders and
+> the `ResourceOffer` diff/blame family are named deferrals). The Git
+> affordance list above has nothing to attach to yet.
+
+**That sentence is literally false about the tree, and revision 1
+repeated it without checking.** `tests/fixtures/pmacs-magit/` is a
+tracked, installable package — 1,914 lines across four modules, with
+`status.lua` spawning git through `pmacs.process.spawn` and parsing
+**`--porcelain=v2 --branch`** into structured sections, plus a 662-line
+acceptance suite (`tests/m8_6_acceptance.rs`, 32 tests) covering
+status, refresh, staging, commit, push and branch behaviour.
+
+**The PRODUCT gap is real and unchanged** — none of that is bundled
+runtime, so a user who installs pmacs gets no git integration. But
+"nothing to attach to" understates what exists to *learn from*, and
+§15's wording should be corrected when this lands.
+
+For a **daily driver**, this is the largest remaining gap. Not because
+git is the most architecturally interesting thing missing — §7
+workspaces and §9 worker identity are both deeper — but because it is
+the one a user touches *every working hour*, and pmacs currently makes
+them leave the editor to answer "what have I changed?".
+
+That is the criterion this lane is chosen against: **frequency of use
+per day**, not depth of model.
+
+## 2. Ground truth — what already exists
+
+Scouted, not assumed:
+
+- **`ProjectKind::Git` is NOT general repository detection**, and
+  revision 1 said it was. Its doc comment is explicit: *"A bare git
+  repository (no language marker found inside)"* (`src/project.rs:89`).
+  Markers are ordered and a language marker beside `.git` **wins**
+  (`src/project.rs:10`), so a normal Rust repository reports
+  `kind = "rust"` and would have been invisible to a lane that gated on
+  `kind == "git"`. That gate would have failed on this very repository.
+
+  **The rule this lane uses instead: never ask pmacs whether it is a
+  git repo.** Run git in the **active file's directory** and let git
+  resolve its own worktree — `git -C <dir> rev-parse --show-toplevel`
+  establishes the root, and a non-zero exit *is* the "not a repository"
+  answer. Git's own resolution handles submodules, worktrees, `GIT_DIR`
+  and `.git` files; a marker walk reimplements a subset of that and
+  gets it subtly wrong.
+- **`pmacs.process.spawn` / `events_take` / `terminate` / `forget`**
+  is the working model for running an external tool asynchronously;
+  `builtin/runtime/compile.lua` is a full worked example, including
+  spawn-failure handling and exit markers.
+- **`pmacs.listview.open`** is a real primitive with existing adopters
+  (`*references*`, `*lsp*`), carrying optional `depth`/`id`,
+  primitive-owned collapse, and selection re-seated by id. `COHERENCE.md`
+  P5 says the remaining work there is **adoption, not construction** —
+  a `*git-status*` panel is exactly that.
+- **Gutter signs exist in both frontends** — the TUI's leading-column
+  glyph (`src/diag.rs`) and the GPU's `GUTTER_SIGN_X` bars
+  (`pmacs-gpu/src/main.rs:420`).
+- **A tested porcelain-v2 parser exists as a package fixture** (above).
+  Its `status.lua` deliberately separates **pure `parse_*` functions
+  that take a string and return structure** from the spawning around
+  them — which is the shape that makes a parser testable without a
+  repository, and it is already proven by 32 tests.
+
+And the constraint that shapes the staging:
+
+- **`DecorationKind` is a CLOSED enum on the wire**
+  (`pmacs-protocol/src/message.rs:1472`): four diagnostic severities,
+  `Selection`, `SearchMatch`, `SearchMatchActive`, `CurrentLine`.
+  **Gutter markers for git hunks therefore require new variants, which
+  is a protocol version bump.** The gutter signs that exist are keyed
+  on `diagnostic_severity_rank` and have no notion of anything else.
+
+## 3. The staging, and why the line falls where it does
+
+**Stage 1 (this lane): read-only, panel-based, NO WIRE CHANGE.**
+
+- `*git-status*` — a `listview` panel over
+  `git status --porcelain=v2 --branch -z` (Q#G-6), rows visiting the
+  file at RET, refreshed by `g` under the completion model in Q#G-1.
+- `*git-diff*` — the diff for the **file** under point (Q#G-7), in a
+  generated buffer rendered as **plain text** (no `diff` grammar
+  exists). **No hunk model** — hunks are Stage 2's concern.
+
+**Stage 2 (separate lane): gutter markers.** Needs new
+`DecorationKind` variants and a `PROTOCOL_VERSION` bump, plus both
+frontends' gutter renderers learning a second rider family.
+
+**Stage 3+ (unscheduled): staging, commit, blame.** Staging and commit
+are where an editor becomes a git *client*; blame is a lower-frequency
+read. Neither belongs in front of the two above.
+
+**The line is drawn at the wire on purpose, and it is a scheduling
+decision as much as a design one.** Parallel lanes are about to start,
+and `PROTOCOL_VERSION` is a strict serialization point — two lanes
+bumping it collide, and this session already recorded what that costs
+(eight broken version assertions on CI from a single bump). Stage 1
+touching no wire is what lets it run **concurrently** with other work.
+Stage 2 must be scheduled alone.
+
+## 4. Coherence impact (§20)
+
+Required by `CLAUDE.md` for coherence-affecting work, and this is
+coherence-affecting — it is §15's named gap.
+
+- **Journey steps touched:** none directly. Git is not currently a
+  journey step; the golden journey runs open → edit → build → test →
+  navigate. This lane does **not** add a step, and I would rather say
+  so than inflate the claim.
+- **§15 contextual affordances — the direct target.** The audit's git
+  affordance list ("a Git change stage/revert/diff") has *nothing to
+  attach to*. Stage 1 creates the thing to attach to; the affordances
+  themselves follow it, and the menu's context vocabulary
+  (`src/menu.rs:44`) would need a `git` context to host them — **out of
+  scope here**, named so it is not forgotten.
+- **§14 workbench primitives — adoption, which is the stated need.**
+  `*git-status*` becomes the **fifth** `listview` call site and the
+  first outside the LSP panels, which is the concrete evidence P5 asks
+  for that the primitive generalizes past its first consumer.
+- **Interaction islands (§6): none added, and this is a real
+  constraint.** The panel gets no hardcoded key interception; it uses
+  `listview`'s existing key handling. §6 records six such shadows and
+  calls them "weak, and growing" — this lane must not make it seven.
+- **Config registry adoption:** at least one setting
+  (`git.enabled`, Q#G-4), defined through `pmacs.config.define` like
+  `ui.line-wrap` and the zoom settings, not a bare Lua global.
+- **Background-work attribution (§9): NEGATIVE, and named as such.**
+  Git runs as a spawned process, and spawned processes do **not** appear
+  in `*workers*` — that view is `async.lua`'s job list; processes live
+  under `pmacs.process.list` (Q#G-5). This lane therefore adds a fifth
+  thing running in the background with no single place to see it. The
+  process is labelled honestly, which is better than anonymous, but
+  **a label is not attribution and this document does not pretend
+  otherwise.** Accepted because these are short-lived reads; it would
+  not be acceptable for Stage 3's push/pull.
+
+## 5. Open questions
+
+### Q#G-1 — is the status panel a snapshot or a live view?
+
+A snapshot is a command that opens a panel; a live view refreshes on
+buffer save, on focus, or on a filesystem watch.
+
+*My vote: **snapshot, refreshed explicitly***, with `g` re-running
+inside the panel. Live refresh needs a watch mechanism, an invalidation
+rule, and a §9 story for the recurring work — all real arcs. A snapshot
+is honest, useful the first day, and does not pretend to a currency it
+cannot maintain.
+
+**But "explicit refresh" does not fit `listview` unmodified, and
+revision 1 missed that.** `listview.refresh` is synchronous:
+
+```lua
+local rows = check_ids(p.on_refresh() or {})   -- listview.lua:402
+```
+
+The result is consumed immediately. `pmacs.process.spawn` cannot return
+rows there — it returns a process id whose output is drained later. So
+revision 1's "adopt `listview`" would have produced exactly one of the
+two failures the reviewer named: a reimplemented list, or a `g` that
+silently does nothing. The primitive's own docs already call a dead `g`
+out as a defect it must not repeat (`listview.lua:416`).
+
+**The completion model, specified.** `on_refresh` stays synchronous and
+honest:
+
+1. **`on_refresh` returns the CURRENT rows immediately**, with a
+   `refreshing…` marker row appended, and *kicks off* the spawn. `g` is
+   therefore never a no-op — it always re-renders and always shows that
+   work started.
+2. **On exit, the completion handler re-opens the panel** via
+   `listview.open` with the same `name` — **and re-seats the selection
+   itself.**
+
+   Revision 2 credited that to the primitive and was wrong.
+   `listview.open` **resets collapse** (`p.collapsed = {}`) and
+   **always seats line 1** (`seat_cursor(p, 1)`,
+   `builtin/runtime/listview.lua:337-378`). The `listview.lua:82` note
+   I cited is about **name** disambiguation to `<2>`, not selection.
+   Only `listview.refresh` preserves a selection, and that is the
+   synchronous path this model cannot use.
+
+   So the contract is explicit and owned here: **capture the selected
+   row's git id (its current path) before re-opening, and after
+   re-opening move to the line whose row carries that id**, computed
+   from the handler's own rows array via `pmacs.editor.move_to_line`.
+   If the id is gone from the new status — the commonest case, since a
+   file that stopped being modified drops out — seat line 1 and say
+   nothing; that is the correct answer, not a failure.
+
+   **Collapse state is moot in Stage 1** because the rows are flat: no
+   `depth`, so nothing to collapse. Stage 2 or a sectioned view would
+   have to revisit this, and would then face the same reset.
+3. **Concurrent refresh is suppressed by a generation counter.** A
+   second `g` while one is in flight bumps the generation; the older
+   completion sees a stale generation and **discards its rows** rather
+   than racing. It does not terminate the first process — reaping is
+   `pmacs.process.forget`'s job and killing git mid-read buys nothing.
+4. **Failure is a row, not a silence.** Non-zero exit or spawn failure
+   renders a row carrying the exit code and the first stderr line, plus
+   a status message. §1.2's silence asymmetry.
+5. **Panel lifetime.** If the panel's buffer is gone when the process
+   exits, the handler drops the result. `compile.lua:252` already
+   handles the buffer-killed case for its own slot; the same shape.
+
+**The alternative — extending `listview` with an async contract — is
+the more correct long-term answer** and is deliberately not taken here:
+it changes a primitive with four existing adopters, and doing that from
+inside its fifth adopter's lane is how a primitive acquires a consumer's
+idiosyncrasies. **If review prefers it, it belongs in its own lane
+before this one.**
+
+### Q#G-0 — what is the relationship to `pmacs-magit`? **(new in rev 2)**
+
+The reviewer's framing of the choice is right: adopt, replace, or
+declare it out-of-product precedent. Doing none of those and quietly
+writing a second parser is the option that must not happen.
+
+*My vote: **port its pure `parse_*` functions and its test corpus into
+the bundled runtime; leave the fixture itself untouched.***
+
+- **The record TOKENIZER is deliberately rewritten, not ported.** The
+  fixture parses **newline-delimited** v2; Stage 1 reads **`-z`**, and
+  those are different grammars — under `-z` a record's fields are
+  NUL-terminated and a rename carries its two paths as separate fields
+  rather than tab-joined. Saying "port the parser" would have been
+  wrong; what ports is the **separation** (pure `parse_*` functions
+  over a string, testable with no repository) and the **case coverage**
+  its 32 tests encode. The tokenizer underneath is new, and its
+  correctness rests on this lane's own corpus.
+- **Port, not import.** The fixture's purpose is to prove the *package
+  system* can host this. If bundled code became its dependency, it
+  would stop demonstrating an independent package and `m8_6` would test
+  less than it claims.
+- **The duplication is therefore deliberate**, and it is the one place
+  this framing accepts two copies of a rule after a session spent
+  removing them. The justification is that they answer different
+  questions — one is product behaviour, one is package-system
+  capability — and coupling them weakens the second. **If review
+  prefers the coupling, that is a defensible call and I will take it**;
+  what I will not do is leave the duplication unstated.
+- **It also settles Q#G-2's format**: the existing, tested parser is
+  **porcelain v2**, so Stage 1 is v2. Revision 1 said v1 for no reason
+  beyond familiarity.
+
+### Q#G-2 — `git` the binary, or a library?
+
+*My vote: **the binary**, via `pmacs.process.spawn`. `compile.lua` is
+the worked precedent, the daemon already spawns external tools, and a
+git library is a dependency with a much larger surface than "run one
+command and parse porcelain". `--porcelain=v2` is explicitly a stable
+machine format; that is what it is for.
+
+**Named risk:** no `git` on `PATH`. §1.2's *silence asymmetry* says the
+failure must be **surfaced with guidance**, not swallowed — the same
+lesson #204 landed for a missing language server.
+
+### Q#G-6 — the status data contract **(new in rev 2)**
+
+Revision 1 said "`--porcelain=v1`" and proposed "a path with a space"
+as the parsing witness. **Both were inadequate.** Porcelain without
+`-z` emits paths in git's **C quoting** for anything non-ASCII or
+containing special characters, and rename/copy records carry *two*
+paths whose separation is positional. A single space-in-path fixture
+proves none of that.
+
+*My vote: the exact invocation*
+
+```
+git --no-optional-locks -C <dir> status --porcelain=v2 --branch -z
+```
+
+**`--no-optional-locks` is part of the contract, not a nicety.**
+`git status` is **not strictly read-only**: it may refresh and write
+the index, and git's own documentation recommends this flag for
+background scripts precisely so a background reader does not contend
+for `index.lock` with the user's real git commands
+(<https://git-scm.com/docs/git-status>). This lane runs status
+*asynchronously, from an editor, while the user may be running git in a
+terminal* — the exact scenario the flag exists for. Revision 2 called
+the lane "read-only" and that was wrong about the mechanism.
+
+It is **witnessed structurally** — the assembled argv is asserted to
+carry the flag — because observing a lock that was *not* taken is not
+something a test can do directly. Verified accepted by the git in use
+here.
+
+The rest: `--porcelain=v2 --branch -z`, also verified accepted. NUL delimiting removes C quoting from
+the problem **entirely** rather than obliging a hand-written unquoter,
+and it makes the two-path rename record unambiguous: the paths are
+separate NUL-terminated fields rather than tab-joined inside one.
+
+The rename/copy identity rule to pin: a `2` record carries the current
+path **and** its origin, and the panel must show which file it is now
+while remembering where it came from — a row whose id is the current
+path, since that is what RET visits.
+
+**Witness corpus, not one case:** modified, added, deleted, untracked,
+**renamed (both paths)**, **copied**, a path with a space, a path with
+a newline, and a non-UTF-8 path. The last two are exactly what `-z`
+buys and what a quoted parser gets wrong.
+
+### Q#G-7 — the diff gesture **(new in rev 2)**
+
+Revision 1 wrote "the diff for the file or hunk under point" while also
+committing RET to visiting the file. **RET cannot do both, there is no
+second binding proposed, and no hunk model exists anywhere in the
+tree.**
+
+*My vote:*
+
+- **RET visits the file** — unchanged, and the behaviour a list of
+  files should have.
+- **A named command, `git.diff-file`, bound to `d` inside the panel.**
+
+  **`d` is not on `listview`'s key surface**, and revision 2 said it
+  was. The bound set is exactly `RET SPC n <down> p <up> TAB g q`
+  (`builtin/runtime/listview.lua:266-279`), bound buffer-locally inside
+  the primitive, which is the only place the panel's buffer handle is
+  known. **Looking the buffer up by name from outside is unsafe** —
+  `listview` deliberately disambiguates a collision to `<2>`, so the
+  name a consumer passed is not necessarily the buffer it got.
+
+  *My vote: **a `keys` table on the open spec***, e.g.
+  `keys = { d = "git.diff-file" }`, bound through the same
+  `bind_local_keymap` that already binds the fixed set. It is additive,
+  general to any adopter, keeps binding where the buffer is known, and
+  adds **no** interception — the §6 constraint holds.
+
+  **The registration lifecycle, which revision 3 omitted and which
+  would have broken the refresh path it depends on.** `Keymap::bind`
+  **refuses duplicates** — `KeymapError::DuplicateBinding`, *"Refuse
+  rather than silently overwrite"* (`src/keymap_tree.rs:75`) — and the
+  completion model calls `listview.open` again on **every** refresh. A
+  naive `keys` implementation therefore errors on the second open, so
+  **every successful refresh would have failed while re-binding `d`.**
+
+  The contract:
+
+  1. **Keys are installed once, when the panel's buffer is created**,
+     and stored on the panel.
+  2. **A later `open` for a live panel does not re-bind.** It
+     **compares** the supplied `keys` against the stored table and
+     **errors on divergence** rather than ignoring it. Silently keeping
+     the old binding would give the consumer a key that does something
+     other than what it just asked for — a dead or lying key, which is
+     the defect `listview` already condemns for `g`.
+  3. **Collisions are rejected at install time**, against both the
+     fixed set (`RET SPC n <down> p <up> TAB g q`) and any
+     **prefix conflict** — `Keymap` has a separate error for turning a
+     leaf into a submap, and a `keys` table must not be able to reach
+     it.
+
+  (The alternative, idempotent re-registration, is tolerable but
+  strictly weaker: it makes a consumer that changes its keys mid-session
+  silently wrong instead of loudly wrong.)
+
+  **This IS a `listview` modification, and revision 2's "no listview
+  modification" was false.** I distinguish it from the async-contract
+  change I deferred: that one alters *when* an existing callback's
+  result is consumed for four existing adopters; this adds an optional
+  field that changes nothing for a spec that omits it. **If review
+  judges any primitive change out of an adopter's lane, the alternative
+  is `listview.open` returning the panel buffer** so the consumer binds
+  its own key — smaller still, but it pushes binding to every adopter.
+- **No hunk model in Stage 1.** Hunks are precisely what gutter markers
+  need, and that is Stage 2's protocol work. Introducing a half hunk
+  model here to serve one gesture would prejudge Stage 2's design from
+  the wrong side.
+
+**And what `d` actually SHOWS, which revision 2 left unstated.** "File,
+not hunk" is a scope, not a contract. A porcelain-v2 row carries an
+**XY** pair — X staged (index vs HEAD), Y unstaged (worktree vs index)
+— and the three plausible diffs answer three different questions:
+`git diff` shows only Y, `--cached` only X, and neither shows an
+untracked file at all.
+
+*My vote: **`d` answers the lane's own question — "what have I
+changed?" — against `HEAD`:***
+
+| row | `d` runs | why |
+|---|---|---|
+| staged, unstaged, or both | `git diff HEAD -- <path>` | one view of the total change; splitting X from Y is a staging UI, which is Stage 3 |
+| deleted | `git diff HEAD -- <path>` | shows the deletion; no special case needed |
+| renamed / copied | `git diff HEAD -- <orig> <current>` | v2 gives both paths; passing both is what lets rename detection render it as a rename rather than an unrelated add+delete |
+| **untracked** | `git diff --no-index -- /dev/null <path>` | **a normal diff shows nothing at all** for an untracked file. Without this case `d` is silently dead on the rows a user is most likely to press it on |
+| non-UTF-8 path | *refuses, with a message* | see Q#G-8 |
+
+The `HEAD` choice is deliberate and is the one thing here I would most
+expect review to push back on: it is the right default for *reading*
+what changed, and the wrong one for *staging*, which is why it is
+correct for Stage 1 and will need revisiting when Stage 3 arrives.
+
+**The exit-state contract, which revision 3 got wrong in two ways.**
+"Non-zero exit renders a failure row" is not correct for `git diff`.
+Both cases below were measured in a scratch repository, not inferred:
+
+**(a) `--no-index` implies `--exit-code`.** It exits **1 when it
+successfully finds differences** — measured: `exit=1` for an untracked
+file against `/dev/null`. Under revision 3's predicate, *every*
+untracked diff — the case `--no-index` exists to serve — would have
+rendered a failure row instead of the diff it just produced.
+
+So for the untracked path the success predicate is **exit ∈ {0, 1}**,
+rendering whatever came out; **exit ≥ 2 is a real failure**. That
+asymmetry is confined to the `--no-index` invocation and does not leak
+to the others, where non-zero still means failure.
+
+**(b) An unborn repository has no `HEAD`.** Measured:
+`git diff HEAD -- <path>` exits **128** with `fatal: bad revision
+'HEAD'`. This is not an edge case — it is a freshly `git init`-ed
+repository with the first files staged, which is exactly when someone
+opens a status panel to see what they are about to commit.
+
+*Policy: **detect once, then split**.*
+
+**Detection needs no extra subprocess.** `--branch` already reports
+`# branch.oid (initial)` when `HEAD` is unborn — observed in the
+output this lane already parses. Revision 4 proposed a separate
+`git rev-parse --verify --quiet HEAD`; that is a second process for a
+fact the first one hands over.
+
+**The reachable states, enumerated from a real unborn repository** —
+`git init`, stage three files, then edit one, delete one, and `git mv`
+one:
+
+```
+# branch.oid (initial)
+1 AD ... ad.txt
+1 AM ... am.txt
+1 A. ... r_new.txt      <- the `git mv`
+? untracked.txt
+```
+
+Two findings fall straight out:
+
+- **`AM` and `AD` are ordinary and carry BOTH states**, which is
+  exactly the gap: `--cached` alone loses the worktree delta, plain
+  `git diff` alone loses the staged base.
+- **Rename and copy CANNOT occur under an unborn `HEAD`.** The
+  `git mv` produced `1 A. … r_new.txt` — an ordinary add of the new
+  path, **not** a `2` record. With no `HEAD` there is nothing to
+  rename *from*, so the rename/copy row class is unreachable here and
+  needs no unborn policy. That is a case closed by evidence rather than
+  handled speculatively.
+
+| unborn row | `d` renders |
+|---|---|
+| `A.` staged only | one patch: `git diff --cached -- <path>` |
+| **`AM` staged + edited** | **two labelled patches** — *staged* `git diff --cached -- <path>`, then *unstaged* `git diff -- <path>` |
+| **`AD` staged + deleted** | **two labelled patches**, same pair; the second renders the deletion |
+| `.M` / `.D` unstaged only | one patch: `git diff -- <path>` |
+| `?` untracked | `git diff --no-index -- /dev/null <path>` (exit ∈ {0,1}) |
+| rename / copy | **unreachable** — see above |
+
+All four `--cached` / plain invocations above were run against that
+repository and render the expected patches.
+
+**The split is unborn-only, and that asymmetry is deliberate.** Once
+`HEAD` exists, `git diff HEAD` gives one total — which is the lane's
+question — and splitting it would be a staging UI (Stage 3). The split
+appears here only because there is no `HEAD` to total *against*.
+
+The generated buffer carries a **header naming what it is showing**:
+*"no commits yet — split view: staged (index) above, unstaged
+(worktree) below"*. Revision 4's wording ("showing staged changes")
+would have described a single total-against-`HEAD` diff, which is
+precisely what this is not. A diff that silently answers a different
+question than the one asked is worse than one that says so — and a
+header that misdescribes a split view is the same failure in smaller
+type.
+
+### Q#G-8 — non-UTF-8 paths: an honest boundary **(new in rev 3)**
+
+Revision 2 listed a non-UTF-8 path in the witness corpus as though it
+were an end-to-end case. **It cannot be**, and the boundary is in the
+bindings: `pmacs.process.spawn` takes `args: Vec<String>`
+(`src/lua_bindings/mod.rs:8683`) and `pmacs.buffer.find_or_open` takes
+`path: String` (`:3564`). Both are Rust `String`, i.e. UTF-8 by
+construction. A path that is valid bytes but not valid UTF-8 can be
+*read* from git's `-z` output and *displayed*, but it cannot be passed
+back to `spawn` for a diff, nor opened.
+
+*My vote: **parse it, show it, and refuse the gesture with a
+message***:
+
+- the row **appears** in the panel, so the user is not lied to about
+  what is modified;
+- **RET and `d` on that row report** that the path is not representable
+  and do nothing else — a witnessed refusal, not a stack trace or a
+  silent no-op;
+- **it is removed from the end-to-end promise.** The witness is
+  parser-and-display **plus the refusal**, and the framing does not
+  claim visiting works.
+
+Making it work end-to-end means `OsString`/bytes through two binding
+boundaries — a real change to the Lua API surface, and not this lane's.
+
+### Q#G-3 — what does the diff view render into?
+
+*My vote: **a generated buffer**, reusing the generated-buffer
+immutability work (Stage 1 merged; that lane's Stage 2 is queued).
+Diff output is read-only text and that machinery exists.
+
+**RESOLVED in rev 2 — there is no bundled `diff` grammar.**
+`BUILTIN_LANGUAGES` (`src/syntax.rs`) has no `diff` entry; checked, not
+assumed. **Stage 1 renders plain generated text**, and diff
+highlighting is later work needing a grammar first.
+
+### Q#G-4 — what is configurable?
+
+*My vote: **one setting to start** — `git.enabled` (boolean, default
+`true`), through the config registry. Resist more until there is use
+evidence; §11's grade is "partial (foundation only)" and adding five
+speculative settings is how a registry becomes noise.
+
+### Q#G-5 — §9 attribution — **RESOLVED, and the answer is negative**
+
+Revision 1 deferred this to implementation. That was wrong: it is
+answerable by reading, and deferring it would have meant discovering a
+known coherence cost *after* committing to the design.
+
+**A spawned git process does not appear in `*workers*` at all.** That
+buffer is `builtin/runtime/async.lua`'s (`:490`) and lists **async
+jobs**; spawned processes live separately under `pmacs.process.list`.
+They are two of the four disjoint activity views §9 grades as
+"mechanism without identity".
+
+So, stated plainly rather than dressed up:
+
+- **This lane adds a fifth thing that runs in the background and is not
+  attributable from one place.** That is a **negative** coherence impact
+  against §9, and it is the honest cost of shipping git status before
+  worker identity exists.
+- **Labelling the process is still required** — a clear label under
+  `pmacs.process.list` is strictly better than an anonymous `git`. But
+  **a label does not solve attribution**, and this document does not
+  claim it does. The claim is only: do not make it worse than it has to
+  be.
+- **The mitigation is bounded in time, not in kind.** These are
+  short-lived reads, not long-running jobs; a `git status` that has not
+  finished is a bug, not a background task a user needs to supervise.
+  That is why the cost is acceptable *now* and would not be for
+  Stage 3's push/pull.
+
+## 6. Verification
+
+- **Parsing, against a corpus rather than a case (Q#G-6):** modified,
+  added, deleted, untracked, **renamed with both paths**, **copied**, a
+  path with a space, and **a path with a newline** — the last is what
+  `-z` buys, and a parser that passes only the space case is the one
+  that ships broken.
+- **A non-UTF-8 path is parsed and displayed, and its gestures refuse
+  with a message** (Q#G-8) — a witnessed refusal at the binding
+  boundary, **not** an end-to-end visit.
+- **The argv carries `--no-optional-locks`** (Q#G-6), asserted
+  structurally. A lock not taken cannot be observed directly, so the
+  invocation is what gets pinned.
+- **`d` is witnessed on every row class** (Q#G-7): staged, unstaged,
+  both, deleted, renamed, and **untracked** — the last because a normal
+  `git diff` shows nothing there, so a missing `--no-index` case makes
+  `d` silently dead exactly where it is most used.
+- **A copy is reported as a COPY, not a rename** (Q#G-7). Porcelain v2
+  folds both into the one `2` record, so `kind` stays `"rename"` for
+  both — every *behaviour* keyed on it is the same — and the
+  distinction is made where it is a distinction: the diff header reads
+  the `<Xscore>` field's leading `R`/`C` and says which one happened.
+  The status row is left alone, because its `XY` prefix already reads
+  `R.` against `C.`. Both classes are asserted, and so is the **argv**:
+  the two-path `git diff HEAD -- <orig> <current>` is right for a copy
+  and a rename alike, so a fix to what the user is *told* must not
+  reach what runs. **Parser-level, deliberately** — the copy ROW is
+  supplied through `_deliver_status` while the repository, the panel,
+  the `d` dispatch and the spawned diff around it are real.
+
+  **The reason, narrowed after review.** This bullet used to say real
+  `git` emits no `2 C` record "even under `status.renames=copies`".
+  **That is too strong, and git's own documentation contradicts it** —
+  `git-status(1)` lists `C` as *"copied (if config option
+  status.renames is set to `copies`)"*. What the test measures is
+  narrower: **for its fixture, whose copy source is left unchanged**,
+  git reports `1 A.`. That is a fact about the fixture, and it is
+  sufficient reason to craft the row — a weaker and true justification
+  in place of a stronger false one. No mechanism is claimed for why an
+  unchanged source is not offered as a candidate; that was never
+  established.
+- **The untracked diff renders on exit 1**, not a failure row (Q#G-7a)
+  — the case `--exit-code` semantics would otherwise break, and the
+  one most likely to be "fixed" later by someone who reads exit 1 as an
+  error.
+- **An unborn repository is witnessed end to end**, and the fixture is
+  **`AM`** specifically — staged then edited again, the shape a first
+  commit actually has partway through. `git init`, stage, edit, open
+  the panel, press `d`, and get **two labelled patches** with the
+  split-view header — not `fatal: bad revision 'HEAD'`, and not a
+  single `--cached` patch that silently drops the worktree edit.
+  **`AD` rides the same fixture**, since one repository can hold both.
+- **Unborn detection reads `# branch.oid (initial)`** from the status
+  output already being parsed — asserted, so nobody later reintroduces
+  a second `rev-parse` process for a fact already in hand.
+- **Rename/copy under an unborn `HEAD` is asserted UNREACHABLE**: the
+  fixture `git mv`s a staged-but-uncommitted file and the parser sees a
+  `1 A.` record, never a `2`. Pinned so a future reader does not
+  "fix" the missing unborn rename policy by inventing one.
+- **Re-binding across a refresh does not error** (Q#G-7): two
+  successive refreshes on a live panel, asserting `d` still works and
+  no `DuplicateBinding` surfaced. This is the one that would have
+  broken on every refresh.
+- **A `keys` table colliding with the fixed set is rejected at install
+  time**, as is a prefix conflict.
+- **Selection is re-seated by the completion handler** (Q#G-1), across
+  a refresh that reorders rows, and **falls back to line 1 without
+  complaint when the selected path drops out of status** — the common
+  case, not an error.
+- **The pure `parse_*` functions are tested without a repository**,
+  which is the shape `pmacs-magit/status.lua` already proves works and
+  the reason to port that separation rather than invent one.
+- **A repository fixture built with real `git`**, in a tempdir, and
+  **bounded with `set_search_boundary`** — R8 was retired two commits
+  ago and is precisely what happens when a fixture lets project
+  detection escape into the developer's environment.
+- **The root rule is witnessed on a repository whose `ProjectKind` is
+  NOT `Git`** — i.e. an ordinary language project with a `.git` beside
+  its manifest. That is the case revision 1's `kind == "git"` gate
+  would have failed, and this repository is one.
+- **Missing `git` on `PATH` is witnessed**, not assumed (Q#G-2), and
+  surfaces guidance rather than silence.
+- **`g` is never a no-op** (Q#G-1): it re-renders and marks that work
+  started, even mid-flight. A dead `g` is a defect `listview` already
+  names.
+- **Concurrent refresh discards the stale generation** rather than
+  racing — asserted by driving two refreshes and completing them out of
+  order.
+- **Failure renders a row**, carrying exit code and stderr.
+- **The panel is a `listview` adopter**, asserted structurally, so a
+  future re-implementation of list behaviour inside git code fails the
+  test rather than passing review.
+- **No new interaction island** — `d` is bound buffer-locally through
+  `listview`'s own binding path (Q#G-7), not a hardcoded interception.
+  §6 stays at six shadows.
+
+Gates via `scripts/gate --acceptance <the new suite>`.
+
+**What this will NOT prove:** that background git work is attributable
+(Q#G-5 — it is not, by construction), or that the parser handles
+porcelain versions other than v2.
+
+## 7. Not in scope
+
+Gutter markers and any `DecorationKind`/`PROTOCOL_VERSION` change
+(Stage 2 — must be scheduled alone). Staging, commit, push, pull,
+branch operations, merge-conflict resolution. Blame. A `git` context in
+the menu vocabulary. Any git *library* dependency. Live refresh
+(Q#G-1). Fixing §9's worker identity — this lane makes it marginally
+worse and says so (Q#G-5). Any hunk model (Q#G-7). Modifying the
+`listview` primitive to carry an async contract — the better long-term
+answer, but it belongs in its own lane before this one, not inside its
+fifth adopter (Q#G-1). Changing `tests/fixtures/pmacs-magit/` or
+`tests/m8_6_acceptance.rs` (Q#G-0).
+
+**A `listview` change IS in scope after all** (Q#G-7): an optional
+`keys` table on the open spec. Revision 2 said no primitive
+modification; that was false, because `d` cannot be bound from outside
+the primitive safely. The async-contract change stays out.
+
+**One correction this lane should carry when it lands:** `COHERENCE.md`
+§15's "no Git integration anywhere in the tree" is literally false —
+`tests/fixtures/pmacs-magit/` exists. The *product* gap it describes is
+real; the sentence needs narrowing to say so.

docs/keybindings.md

@@ -231,12 +231,33 @@ all share one keymap:
 | `RET` / `SPC` | `listview.visit` — act on the item under the cursor |
 | `n` / `<down>` | `cursor.down` |
 | `p` / `<up>` | `cursor.up` |
+| `TAB` | `listview.toggle` — collapse/expand the tree node under the cursor; a panel with no tree rows delegates to `buffer.tab` |
 | `g` | `listview.refresh` — re-run the data source and re-render |
 | `q` | `listview.quit` — restore the buffer that was active before the panel opened |
 
+(`TAB` arrived with the tree primitive and this table had not recorded
+it. Noted rather than quietly added: the omission predates the git lane
+that found it.)
+
 Panels currently built on this: `*references*`, `*outline*`,
-`*lsp-help*` (hover docs). Header text always spells out the same
+`*lsp-help*` (hover docs), `*lsp*` (`lsp.status`), and `*git-status*`
-`RET`/`n`/`p`/`g`/`q` legend inline.
+(`git.status`). Header text always spells out the panel's own legend
+inline.
+
+A panel may add keys of its own through an optional `keys` table on the
+open spec, bound through the same buffer-local path — so they are
+inspectable by `describe-key` and rebindable from `init.lua`, exactly
+like the fixed set. They are installed once with the panel's buffer and
+may not collide with the fixed set, nor prefix it. One panel uses this
+today:
+
+| Buffer | Key | Command |
+|---|---|---|
+| `*git-status*` (`git.status`) | `d` | `git.diff-file` — the diff for the file under the cursor, into `*git-diff*` |
+
+`git.status` gets **no global chord**: an opening key is a
+command-surface decision the Stage 1 framing did not make, so the entry
+point is `M-x git.status`.
 
 `*buffer-list*` (`editor.list-buffers`, `C-x C-b`) uses its own
 keymap, layered on the same idiom, in `builtin/commands/default.lua`:

docs/lsp-file-watcher-framing.md

@@ -0,0 +1,240 @@
+# LSP file watcher — framing
+
+**Status: revision 2 — approved design (2026-08-10), plus two
+correctness findings from review OF THE IMPLEMENTATION.** The user ruled
+that D1 and D2 proceed with the walking explicitly surviving this lane;
+D3 gets its own framing. The acceptance bar for this lane is correctness
+and the leak, not "the flipping stops".
+
+**Revision 2 records a review round against the code, not the design.
+Both findings are cases where the first fix was itself wrong**, and both
+were confirmed against the tree before being acted on:
+
+- **P1 — the form must be read from the PATTERN, not from the union
+  arm.** `resolve_watcher` returned `"absolute"` for *every* string, so
+  a bare `*.txt` — a valid relative pattern under LSP 3.17, and how VS
+  Code treats string watchers across workspace folders — was matched
+  against `<base>/foo.txt` and could never fire. **That case worked
+  before this lane touched it**, so the repair for #233 silently broke a
+  live path while fixing another. A leading `/` is what makes a pattern
+  absolute; classification now reads the string.
+- **P2 — a scan completing after cancellation still emitted.**
+  `scan_tree` awaits `read_dir` once per directory, so the coroutine
+  sits suspended for most of a tick with `_sleep` already cleared. A
+  cancel arriving there sets `cancelled` and has no sleep to interrupt,
+  and the resumed scan ran on to `did_change_watched_files` — one stale
+  batch under the superseded pattern, which is a wrong-pattern
+  notification the server acts on. Cancellation and liveness are now
+  rechecked after the scan.
+
+**F1's lesson repeated itself inside this lane.** The flat-pattern test
+constrains the RelativePattern **object** arm, so it said nothing about
+the **string** arm P1's regression lived in — the same
+tested-path/exercised-path split this framing opened by naming. Both
+findings now have tests, and both tests were mutation-checked: each
+fails only its own defect.
+
+**A test seam was added, and is recorded here rather than buried.**
+`pmacs.lsp._after_scan_for_tests` is a production hook, nil in normal
+operation, that P2's witness requires: the race is a cancel landing
+during one of the scan's suspensions, which no arrangement of real
+timing produces on demand. Same device and justification as `git.lua`'s
+`_deliver_status`. It is handed the scan result deliberately — a test
+that cancels on any *other* scan passes with the fix deleted, because
+the loop would break at the post-sleep check and emit nothing anyway.
+
+Answers issue #233. **Scope is D1 and D2 only** — the two bug-shaped
+defects. D3 (the polling cost) is named here, deferred with reasons, and
+gets its own framing.
+
+## What is and is not a regression
+
+**#232 is not at fault and nothing about it should be reverted.** The
+statusline activity indicator it added is *correct*: it renders real
+in-flight jobs from `AsyncRuntime::activity_summary`, and the jobs it
+names (`sleep 250ms`, `read_dir <path>`) are real. What changed on
+2026-08-09 is **visibility**, not behaviour.
+
+The behaviour has been there since `1c25730` (2026-05-19). So the user-
+facing report — "the modeline flips several times a second" — is a
+three-month-old defect that became observable last week, and the fix
+belongs to the watcher, not the indicator.
+
+Recorded plainly because the tempting move is to quiet the indicator,
+and that would delete the only instrument that found this.
+
+## Verified against the tree at `0e4c58d`
+
+Every claim below was read or executed this session, not carried from
+the issue.
+
+- `FILE_WATCH_INTERVAL_MS = 250` (`lsp.lua:1924`); each watcher is one
+  `pmacs.async` coroutine looping sleep → `scan_tree`
+  (`lsp.lua:2060-2097`).
+- `scan_tree` builds `rel` from an empty prefix and calls
+  `matches(rel)` — **relative** paths (`lsp.lua:2035-2056`).
+- **`walk` recurses into every directory unconditionally.** `matches`
+  gates only whether an entry is *recorded*. A watcher that can never
+  match still walks the whole tree every tick.
+- `resolve_watcher`'s string branch returns the pattern **unchanged**
+  with the base guessed from an attached file's directory
+  (`lsp.lua:2102-2116`).
+- `register_file_watchers` ends `file_watchers[skey][reg.id] = recs`
+  with no cancellation of the outgoing list (`lsp.lua:2132`).
+- Job purposes are `format!("sleep {}ms", …)` (`async_runtime.rs:1027`)
+  and `format!("read_dir {}", …)` (`:1178`).
+- The fake LSP registers **one** watcher, a `RelativePattern`
+  `{ baseUri, pattern: "**/*.txt" }`, id `watch-1`
+  (`pmacs_fake_lsp.rs:312-331`).
+
+### The glob table, reproduced
+
+Ran the tree's own `expand_braces` / `glob_one_to_pattern` /
+`glob_matcher` under LuaJIT. Output matches the issue exactly, compiled
+patterns included:
+
+| glob | compiled | `main.go` | `go.mod` | absolute |
+|---|---|---|---|---|
+| `**/*.{mod,work}` | `^.-[^/]*%.mod$` | false | **true** | true |
+| `<abs>/goproj/**/*.{go,…}` | `^/tmp/goproj/.-[^/]*%.go$` | false | false | true |
+| `<abs>/rsproj/**/*.rs` | `^/tmp/rsproj/.-[^/]*%.rs$` | false | false | true |
+
+## Two findings the issue does not carry, both of which shape the fix
+
+### F1 — the existing test cannot discriminate this fix, in either direction
+
+`**/*.txt` compiles to `^.-[^/]*%.txt$`, and `.-` spans `/`. Measured:
+it matches `a.txt`, `sub/a.txt`, `/base/a.txt` **and**
+`/base/sub/a.txt`. So `m4_24_workspace_did_change_watched_files` passes
+whether the matching subject is relative or absolute.
+
+The issue says the tested path and the exercised path are disjoint. The
+sharper statement is that the existing test is **insensitive**: it
+cannot fail for D1 and it cannot confirm D1's fix. New coverage must use
+a pattern whose two readings disagree, or it will inherit the same
+blindness.
+
+### F2 — the fix cannot simply "match absolute"; the form must be carried
+
+Per LSP, a plain-string glob matches the **absolute** path while a
+`RelativePattern`'s pattern is relative to **its base**. Matching
+everything absolutely breaks the second. Measured on `*.txt`:
+
+| subject | matches |
+|---|---|
+| `a.txt` (relative, correct for RelativePattern) | **true** |
+| `/base/a.txt` (absolute) | **false** |
+
+`resolve_watcher` returns `(base, pattern)` and **discards which form it
+came from**, so both callers below it are already unable to tell. The
+fix therefore changes that function's contract — a third return value or
+an explicit record field — rather than only changing the subject string
+at the match site. A fix that ignores this trades rust-analyzer's six
+broken globs for every `RelativePattern` whose pattern does not begin
+`**/`.
+
+## D1 — plain-string globs never match
+
+**Consequences, as measured in the issue and confirmed by the table
+above:** rust-analyzer is never told about any file change (all six
+globs absolute); gopls is told about `go.mod`/`go.work` but never `.go`
+sources (only its relative glob matches).
+
+**Fix:** match a plain-string glob against `base .. "/" .. rel`; keep a
+`RelativePattern` matched against `rel`. `resolve_watcher` gains the
+form in its return, and the record carries it.
+
+The leading `**/` in gopls' relative glob compiles to `.-`, which spans
+`/`, so that glob keeps matching under the absolute subject — which is
+why one server's working case does not regress.
+
+## D2 — re-registration leaks the previous coroutines
+
+`file_watchers[skey][reg.id] = recs` replaces the record list without
+setting `cancelled` or cancelling the in-flight `_sleep`. The old
+coroutines poll until the server dies and are unreachable by
+`unregister_file_watchers`, which can only see what the table now holds.
+
+**Reachable today**: rust-analyzer registers
+`workspace/didChangeWatchedFiles` **twice under the same id**, six
+watchers each, with no intervening unregister — 12 concurrent
+coroutines, six permanently uncancellable. The issue's 44.1/s dir-open
+rate against a ~270 ms period implies 12 watchers, so the leak is
+measured from outside the process, not only read from the source.
+
+**Fix:** cancel the outgoing list before replacing it, with the same
+treatment `unregister_file_watchers` already applies.
+
+## What this lane does NOT fix, stated so the report is not mistaken for closed
+
+**The poll cost survives both fixes.** D1 makes matching correct and D2
+halves rust-analyzer's watcher count; neither stops the walk. After this
+lane, rust-analyzer still walks the entire tree every 250 ms — six times
+per tick instead of twelve — including `.git`, `target` and
+`node_modules`, at one async job per directory.
+
+So the modeline will still show activity, at roughly half the rate. **If
+the acceptance bar for this lane is "the flipping stops", this lane does
+not meet it** and should not be started until D3 is framed. That is a
+ruling for the user, not an assumption to make quietly.
+
+**Answered 2026-08-10: the user accepted this scope.** D1 and D2
+proceed; the walking is D3's problem, framed separately.
+
+## D3 — deferred, with what was checked
+
+Options named in the issue: coalesce a server's watchers into one scan;
+root the scan at the workspace rather than an attached file's directory;
+an ignore list; back off when nothing changes; or a real
+filesystem-notification primitive.
+
+Checked while framing: **there is no `notify`/inotify dependency in the
+tree**, so the last option is a new crate *and* a new Rust primitive
+plus its Lua binding — not a small change. There is also **no existing
+ignore-list infrastructure** to reuse; `src/project.rs` knows `.git` as
+a *marker* name, not as something to skip.
+
+D3 is a `COHERENCE.md` §9 concern — background work with no ownership
+model — and §9's own Stage 1 is the indicator that surfaced it.
+
+## Verification
+
+The suite must fail without each fix, which the existing suite cannot
+(F1). Planned:
+
+- **A fake-LSP mode registering a plain-string ABSOLUTE glob**, with a
+  pattern whose relative and absolute readings **disagree** — so the
+  test fails today and passes after D1.
+- **A fake-LSP mode registering a `RelativePattern` whose pattern does
+  not begin `**/`** (e.g. `*.txt` at the base). This is F2's guard: it
+  passes today, and fails against a fix that matches everything
+  absolutely. Without it, the obvious wrong fix is green.
+- **A re-registration mode**: the same id twice, no unregister. The
+  witness is that the superseded watchers **stop**, asserted on
+  observable polling rather than on internal table shape, since the
+  defect is precisely that the old records are unreachable.
+- Existing `m4_24` kept and expected **unchanged** — it covers the
+  working branch and its insensitivity is now recorded rather than
+  mistaken for coverage.
+
+Each new test is mutation-tested against the fix it names.
+
+## Coherence impact (§20)
+
+- **Journey steps**: none added; step 5's editing surface is affected
+  only in that a correct watcher makes servers see edits they currently
+  miss.
+- **Interaction islands**: none.
+- **Config registry**: no new setting. The interval stays a module
+  constant; making it configurable would offer the user a knob for a
+  defect rather than a preference, and D3 may remove the poll entirely.
+- **Background-work attribution (§9)**: this lane *reduces* unattributed
+  background work but does not model it. D3 owns that, and the honest
+  statement is that the indicator worked — it made three months of
+  invisible churn visible on its first week.
+
+## Gates
+
+`./scripts/gate --acceptance m4_acceptance` plus the touched LSP
+acceptance suites; no `--protocol` (no wire change, no
+`PROTOCOL_VERSION` bump).

src/bin/pmacs_fake_lsp.rs

@@ -332,6 +332,106 @@ fn main() {
                 });
                 write_frame(&mut stdout, &req);
             }
+            // Issue #233 D1: `filewatchabs` registers the same watcher
+            // as a PLAIN-STRING glob — `<base>/**/*.txt`, the form
+            // rust-analyzer and gopls actually send. Per LSP it matches
+            // the file's ABSOLUTE path; its relative reading matches
+            // nothing, so the mode discriminates the match subject.
+            ("initialized", _) if mode == "filewatchabs" => {
+                let base = std::env::var("PMACS_FAKE_LSP_WATCH_BASE").unwrap_or_default();
+                let req = serde_json::json!({
+                    "jsonrpc": "2.0",
+                    "id": 9301,
+                    "method": "client/registerCapability",
+                    "params": { "registrations": [{
+                        "id": "watch-abs",
+                        "method": "workspace/didChangeWatchedFiles",
+                        "registerOptions": { "watchers": [{
+                            "globPattern": format!("{base}/**/*.txt"),
+                            "kind": 7
+                        }] }
+                    }] }
+                });
+                write_frame(&mut stdout, &req);
+            }
+            // Issue #233 review P1 guard: `filewatchbare` registers a
+            // BARE STRING with no base and no leading `/` — `*.txt`.
+            // The string arm and the `filewatchflat` arm below carry the
+            // same pattern deliberately: `flat` proves a
+            // RelativePattern stays relative, and this proves the
+            // classification is read from THE PATTERN rather than from
+            // the union arm it arrived in. The first fix for #233
+            // called every string absolute, which matched this against
+            // `<base>/foo.txt` and broke a case that had worked since
+            // May. Without this mode that regression is invisible.
+            ("initialized", _) if mode == "filewatchbare" => {
+                let req = serde_json::json!({
+                    "jsonrpc": "2.0",
+                    "id": 9304,
+                    "method": "client/registerCapability",
+                    "params": { "registrations": [{
+                        "id": "watch-bare",
+                        "method": "workspace/didChangeWatchedFiles",
+                        "registerOptions": { "watchers": [{
+                            "globPattern": "*.txt",
+                            "kind": 7
+                        }] }
+                    }] }
+                });
+                write_frame(&mut stdout, &req);
+            }
+            // Issue #233 F2 guard: `filewatchflat` registers a
+            // RelativePattern whose pattern has no leading `**/`
+            // (`*.txt` at the base). It matches base-level files
+            // RELATIVELY and no absolute path at all, so a fix that
+            // matches every form against the absolute path goes red.
+            ("initialized", _) if mode == "filewatchflat" => {
+                let base = std::env::var("PMACS_FAKE_LSP_WATCH_BASE").unwrap_or_default();
+                let req = serde_json::json!({
+                    "jsonrpc": "2.0",
+                    "id": 9302,
+                    "method": "client/registerCapability",
+                    "params": { "registrations": [{
+                        "id": "watch-flat",
+                        "method": "workspace/didChangeWatchedFiles",
+                        "registerOptions": { "watchers": [{
+                            "globPattern": {
+                                "baseUri": format!("file://{base}"),
+                                "pattern": "*.txt"
+                            },
+                            "kind": 7
+                        }] }
+                    }] }
+                });
+                write_frame(&mut stdout, &req);
+            }
+            // Issue #233 D2: `filewatchrereg` registers the SAME id
+            // twice with no unregister between — `**/*.old` then
+            // `**/*.new` — exactly rust-analyzer's shape. The second
+            // registration must supersede the first: only `.new`
+            // events may ever reach `.received`.
+            ("initialized", _) if mode == "filewatchrereg" => {
+                let base = std::env::var("PMACS_FAKE_LSP_WATCH_BASE").unwrap_or_default();
+                for (rid, pattern) in [(9303, "**/*.old"), (9304, "**/*.new")] {
+                    let req = serde_json::json!({
+                        "jsonrpc": "2.0",
+                        "id": rid,
+                        "method": "client/registerCapability",
+                        "params": { "registrations": [{
+                            "id": "watch-re",
+                            "method": "workspace/didChangeWatchedFiles",
+                            "registerOptions": { "watchers": [{
+                                "globPattern": {
+                                    "baseUri": format!("file://{base}"),
+                                    "pattern": pattern
+                                },
+                                "kind": 7
+                            }] }
+                        }] }
+                    });
+                    write_frame(&mut stdout, &req);
+                }
+            }
             ("initialized", _) => {}
             // T M4.5: the client's file-watch notifications. Append
             // `type uri` lines to `<base>/.received` as a test

src/editor.rs

@@ -797,6 +797,20 @@ impl EditorState {
                 include_str!("../builtin/runtime/linewrap.lua"),
             )
             .expect("load linewrap builtin chunk");
+        // Git integration Stage 1 (docs/git-integration-framing.md):
+        // `*git-status*` and `*git-diff*`. Loaded after `listview.lua`,
+        // whose `open` (and whose new optional `keys` table) it drives,
+        // and after `window.lua`, which owns `window.panel-height` — the
+        // setting a `display = "panel"` listview resolves. It binds no
+        // global key: an opening chord is a command-surface decision and
+        // the framing did not make one, so the entry point is
+        // `M-x git.status`.
+        lua_host
+            .eval(
+                Some("@pmacs/builtin/runtime/git.lua"),
+                include_str!("../builtin/runtime/git.lua"),
+            )
+            .expect("load git builtin chunk");
         // T M7.11 bundled-package bootstrap. Through M7.10 the REPL
         // was loaded directly via `eval(include_str!(...))`; the
         // M7.11 deliverable migrates it to the package system so it

tests/m4_acceptance.rs

@@ -5351,6 +5351,416 @@ fn m4_24_workspace_did_change_watched_files() {
     );
 }
 
+/// Issue #233 D1 — a PLAIN-STRING `GlobPattern` matches the file's
+/// ABSOLUTE path (LSP 3.17), not the walk's relative path. The
+/// `filewatchabs` fake registers `<base>/**/*.txt` as a bare string —
+/// the form rust-analyzer and gopls actually send. Its relative
+/// reading matches nothing (an anchored `^<base>/…` can never match
+/// `foo.txt`), so before the fix no event could ever be reported.
+/// The watcher's base is guessed from the attached file's directory —
+/// the tempdir here, and the production path for bare-string globs.
+#[test]
+fn m4_24_plain_string_glob_matches_absolute_path() {
+    use pmacs::editor::EditorState;
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let base = dir.path().to_path_buf();
+    let base_disp = base.display().to_string();
+    let a_path = base.join("a.rs");
+    std::fs::write(&a_path, b"fn a() {}\n").expect("write a");
+    let a_disp = a_path.display().to_string();
+    let received = base.join(".received");
+    let foo_uri = format!("file://{}", base.join("foo.txt").display());
+
+    let mut state = EditorState::new_with_roots(&crate::iso::roots());
+    let fake = fake_lsp_path();
+    state
+        .lua_host
+        .lua()
+        .load(format!(
+            "pmacs.lsp.config.rust = {{ command = '{fake}',
+               env = {{ PMACS_FAKE_LSP_MODE = 'filewatchabs',
+                        PMACS_FAKE_LSP_WATCH_BASE = '{base_disp}' }} }}"
+        ))
+        .exec()
+        .expect("override rust config");
+    state
+        .lua_host
+        .lua()
+        .load(format!("pmacs.buffer.find_or_open('{a_disp}')"))
+        .exec()
+        .expect("open a.rs");
+    assert!(
+        pump_lua_flag(
+            &mut state,
+            "(function() for _,r in ipairs(pmacs.lsp.list()) do \
+               if r.state and r.state.kind=='initialized' then return true end \
+             end return false end)()",
+            5,
+        ),
+        "fake never initialized"
+    );
+
+    // Same warm-up as m4_24: let registerCapability land and the
+    // watcher take its empty baseline before files appear.
+    let warm = Instant::now() + Duration::from_millis(900);
+    while Instant::now() < warm {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    std::fs::write(base.join("foo.txt"), b"one\n").expect("write foo.txt");
+    std::fs::write(base.join("bar.md"), b"md\n").expect("write bar.md");
+    assert!(
+        pump_until_file_contains(&mut state, &received, &format!("1 {foo_uri}"), 6),
+        "CREATED for foo.txt never reported under a plain-string glob; \
+         .received = {:?}",
+        std::fs::read_to_string(&received).unwrap_or_default()
+    );
+    assert!(
+        !std::fs::read_to_string(&received)
+            .unwrap_or_default()
+            .contains("bar.md"),
+        "non-matching .md must be filtered out"
+    );
+}
+
+/// Issue #233 F2 guard — a `RelativePattern` stays relative to its
+/// base. The `filewatchflat` fake registers `{ baseUri, pattern =
+/// "*.txt" }`, whose pattern has no leading `**/`: it matches
+/// base-level files RELATIVELY and cannot match any absolute path
+/// (`[^/]*` spans no `/`). Green before and after D1's fix; red
+/// against the obvious wrong fix that matches every form absolutely.
+/// `sub/nested.txt` pins the other half of the same contract: a
+/// base-level pattern must not match into subdirectories.
+#[test]
+fn m4_24_relative_pattern_without_globstar_stays_relative() {
+    use pmacs::editor::EditorState;
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let base = dir.path().to_path_buf();
+    let base_disp = base.display().to_string();
+    let a_path = base.join("a.rs");
+    std::fs::write(&a_path, b"fn a() {}\n").expect("write a");
+    let a_disp = a_path.display().to_string();
+    let received = base.join(".received");
+    let foo_uri = format!("file://{}", base.join("foo.txt").display());
+    std::fs::create_dir(base.join("sub")).expect("mkdir sub");
+
+    let mut state = EditorState::new_with_roots(&crate::iso::roots());
+    let fake = fake_lsp_path();
+    state
+        .lua_host
+        .lua()
+        .load(format!(
+            "pmacs.lsp.config.rust = {{ command = '{fake}',
+               env = {{ PMACS_FAKE_LSP_MODE = 'filewatchflat',
+                        PMACS_FAKE_LSP_WATCH_BASE = '{base_disp}' }} }}"
+        ))
+        .exec()
+        .expect("override rust config");
+    state
+        .lua_host
+        .lua()
+        .load(format!("pmacs.buffer.find_or_open('{a_disp}')"))
+        .exec()
+        .expect("open a.rs");
+    assert!(
+        pump_lua_flag(
+            &mut state,
+            "(function() for _,r in ipairs(pmacs.lsp.list()) do \
+               if r.state and r.state.kind=='initialized' then return true end \
+             end return false end)()",
+            5,
+        ),
+        "fake never initialized"
+    );
+
+    let warm = Instant::now() + Duration::from_millis(900);
+    while Instant::now() < warm {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    // nested.txt is written BEFORE foo.txt, so a watcher that wrongly
+    // matched it would report it no later than foo.txt's event — the
+    // negative assertion after the positive one is race-free.
+    std::fs::write(base.join("sub").join("nested.txt"), b"deep\n").expect("write nested.txt");
+    std::fs::write(base.join("foo.txt"), b"one\n").expect("write foo.txt");
+    assert!(
+        pump_until_file_contains(&mut state, &received, &format!("1 {foo_uri}"), 6),
+        "CREATED for base-level foo.txt never reported under a \
+         RelativePattern without `**/`; .received = {:?}",
+        std::fs::read_to_string(&received).unwrap_or_default()
+    );
+    assert!(
+        !std::fs::read_to_string(&received)
+            .unwrap_or_default()
+            .contains("nested.txt"),
+        "a base-level `*.txt` RelativePattern must not match into \
+         subdirectories"
+    );
+}
+
+/// Issue #233 review P2 — a scan that completes AFTER cancellation
+/// must not emit.
+///
+/// `scan_tree` awaits `read_dir` once per directory, so the watcher
+/// coroutine spends most of a tick suspended with `_sleep` already
+/// cleared. A cancel arriving there — re-registration or unregistration
+/// — sets `cancelled` and has no sleep to interrupt, so before the fix
+/// the resumed scan ran on and emitted one last batch under the
+/// superseded pattern.
+///
+/// No arrangement of real timing produces that interleaving on demand,
+/// so it is driven through `pmacs.lsp._after_scan_for_tests`, the same
+/// device `git.lua` uses for out-of-order completions. The hook is
+/// handed the scan result and cancels **only on the scan that observed
+/// `foo.txt`** — cancelling on any other scan would pass with the fix
+/// deleted, because the loop would break at the post-sleep check and
+/// emit nothing regardless.
+#[test]
+fn m4_24_a_scan_finishing_after_cancellation_emits_nothing() {
+    use pmacs::editor::EditorState;
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let base = dir.path().to_path_buf();
+    let base_disp = base.display().to_string();
+    let a_path = base.join("a.rs");
+    std::fs::write(&a_path, b"fn a() {}\n").expect("write a");
+    let a_disp = a_path.display().to_string();
+    let received = base.join(".received");
+
+    let mut state = EditorState::new_with_roots(&crate::iso::roots());
+    let fake = fake_lsp_path();
+    state
+        .lua_host
+        .lua()
+        .load(format!(
+            "pmacs.lsp.config.rust = {{ command = '{fake}',
+               env = {{ PMACS_FAKE_LSP_MODE = 'filewatch',
+                        PMACS_FAKE_LSP_WATCH_BASE = '{base_disp}' }} }}"
+        ))
+        .exec()
+        .expect("override rust config");
+    state
+        .lua_host
+        .lua()
+        .load(format!("pmacs.buffer.find_or_open('{a_disp}')"))
+        .exec()
+        .expect("open a.rs");
+    assert!(
+        pump_lua_flag(
+            &mut state,
+            "(function() for _,r in ipairs(pmacs.lsp.list()) do \
+               if r.state and r.state.kind=='initialized' then return true end \
+             end return false end)()",
+            5,
+        ),
+        "fake never initialized"
+    );
+
+    // Armed BEFORE the file exists, so the cancel cannot land early:
+    // the hook fires on every scan and only cancels once the scan it is
+    // inspecting actually contains foo.txt.
+    state
+        .lua_host
+        .lua()
+        .load(
+            "pmacs.lsp._after_scan_for_tests = function(record, cur)
+               if cur and cur['foo.txt'] then record.cancelled = true end
+             end",
+        )
+        .exec()
+        .expect("install scan hook");
+
+    let warm = Instant::now() + Duration::from_millis(900);
+    while Instant::now() < warm {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    std::fs::write(base.join("foo.txt"), b"one\n").expect("write foo.txt");
+
+    let deadline = Instant::now() + Duration::from_secs(4);
+    while Instant::now() < deadline {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    let got = std::fs::read_to_string(&received).unwrap_or_default();
+    assert!(
+        !got.contains("foo.txt"),
+        "a watcher cancelled during its scan emitted a stale batch \
+         anyway; .received = {got:?}"
+    );
+}
+
+/// Issue #233 review P1 — a BARE-STRING glob with no leading `/` is a
+/// relative pattern and must stay one.
+///
+/// The first fix for #233 classified every string-arm pattern as
+/// absolute, so `*.txt` was matched against `<base>/foo.txt` and could
+/// never fire — silently breaking a case that had worked since May
+/// while fixing the absolute one. `m4_24_relative_pattern_without_globstar_stays_relative`
+/// does not cover it: that mode sends the `RelativePattern` OBJECT form,
+/// so it constrains the object arm only. This sends the same pattern
+/// through the STRING arm, which is the arm the regression lived in.
+#[test]
+fn m4_24_bare_string_glob_stays_relative() {
+    use pmacs::editor::EditorState;
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let base = dir.path().to_path_buf();
+    let base_disp = base.display().to_string();
+    let a_path = base.join("a.rs");
+    std::fs::write(&a_path, b"fn a() {}\n").expect("write a");
+    let a_disp = a_path.display().to_string();
+    let received = base.join(".received");
+    let foo_uri = format!("file://{}", base.join("foo.txt").display());
+
+    let mut state = EditorState::new_with_roots(&crate::iso::roots());
+    let fake = fake_lsp_path();
+    state
+        .lua_host
+        .lua()
+        .load(format!(
+            "pmacs.lsp.config.rust = {{ command = '{fake}',
+               env = {{ PMACS_FAKE_LSP_MODE = 'filewatchbare',
+                        PMACS_FAKE_LSP_WATCH_BASE = '{base_disp}' }} }}"
+        ))
+        .exec()
+        .expect("override rust config");
+    state
+        .lua_host
+        .lua()
+        .load(format!("pmacs.buffer.find_or_open('{a_disp}')"))
+        .exec()
+        .expect("open a.rs");
+    assert!(
+        pump_lua_flag(
+            &mut state,
+            "(function() for _,r in ipairs(pmacs.lsp.list()) do \
+               if r.state and r.state.kind=='initialized' then return true end \
+             end return false end)()",
+            5,
+        ),
+        "fake never initialized"
+    );
+
+    let warm = Instant::now() + Duration::from_millis(900);
+    while Instant::now() < warm {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    std::fs::write(base.join("foo.txt"), b"one\n").expect("write foo.txt");
+    assert!(
+        pump_until_file_contains(&mut state, &received, &format!("1 {foo_uri}"), 6),
+        "CREATED for foo.txt never reported under a bare-string `*.txt` \
+         glob — the string arm is being classified absolute again; \
+         .received = {:?}",
+        std::fs::read_to_string(&received).unwrap_or_default()
+    );
+}
+
+/// Issue #233 D2 — re-registering a live id supersedes it. The
+/// `filewatchrereg` fake registers `watch-re` TWICE with no
+/// unregister between — `**/*.old`, then `**/*.new` — exactly the
+/// shape rust-analyzer sends. The superseded watchers must STOP,
+/// asserted on observable polling rather than on table shape (the
+/// defect is precisely that the replaced records become unreachable
+/// while still polling): `f.old` exists on disk before either `.new`
+/// event lands, so a leaked first-registration watcher, polling at
+/// the same 250 ms cadence, would have reported it by the time the
+/// second `.new` positive arrives.
+#[test]
+fn m4_24_reregistration_supersedes_previous_watchers() {
+    use pmacs::editor::EditorState;
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let base = dir.path().to_path_buf();
+    let base_disp = base.display().to_string();
+    let a_path = base.join("a.rs");
+    std::fs::write(&a_path, b"fn a() {}\n").expect("write a");
+    let a_disp = a_path.display().to_string();
+    let received = base.join(".received");
+    let f_old_uri = format!("file://{}", base.join("f.old").display());
+    let f_new_uri = format!("file://{}", base.join("f.new").display());
+    let g_new_uri = format!("file://{}", base.join("g.new").display());
+
+    let mut state = EditorState::new_with_roots(&crate::iso::roots());
+    let fake = fake_lsp_path();
+    state
+        .lua_host
+        .lua()
+        .load(format!(
+            "pmacs.lsp.config.rust = {{ command = '{fake}',
+               env = {{ PMACS_FAKE_LSP_MODE = 'filewatchrereg',
+                        PMACS_FAKE_LSP_WATCH_BASE = '{base_disp}' }} }}"
+        ))
+        .exec()
+        .expect("override rust config");
+    state
+        .lua_host
+        .lua()
+        .load(format!("pmacs.buffer.find_or_open('{a_disp}')"))
+        .exec()
+        .expect("open a.rs");
+    assert!(
+        pump_lua_flag(
+            &mut state,
+            "(function() for _,r in ipairs(pmacs.lsp.list()) do \
+               if r.state and r.state.kind=='initialized' then return true end \
+             end return false end)()",
+            5,
+        ),
+        "fake never initialized"
+    );
+
+    let warm = Instant::now() + Duration::from_millis(900);
+    while Instant::now() < warm {
+        state.tick_processes();
+        state.tick_lsp();
+        state.tick_async();
+        std::thread::sleep(Duration::from_millis(15));
+    }
+
+    std::fs::write(base.join("f.old"), b"old\n").expect("write f.old");
+    std::fs::write(base.join("f.new"), b"new\n").expect("write f.new");
+    assert!(
+        pump_until_file_contains(&mut state, &received, &format!("1 {f_new_uri}"), 6),
+        "CREATED for f.new never reported by the superseding watcher; \
+         .received = {:?}",
+        std::fs::read_to_string(&received).unwrap_or_default()
+    );
+    // A second positive puts at least one more full poll cycle between
+    // f.old appearing on disk and the negative assertion below.
+    std::fs::write(base.join("g.new"), b"new\n").expect("write g.new");
+    assert!(
+        pump_until_file_contains(&mut state, &received, &format!("1 {g_new_uri}"), 6),
+        "CREATED for g.new never reported by the superseding watcher"
+    );
+    assert!(
+        !std::fs::read_to_string(&received)
+            .unwrap_or_default()
+            .contains(&f_old_uri),
+        "the superseded `**/*.old` watcher is still polling after \
+         re-registration under the same id; .received = {:?}",
+        std::fs::read_to_string(&received).unwrap_or_default()
+    );
+}
+
 /// Tier 1 single-binary language servers ship pre-configured in the
 /// default bundle. Binary-independent: we don't spawn anything, just
 /// assert the `pmacs.lsp.config` tables and the `pmacs.lsp.filetypes`