refactor: organize tests and dry (#49)

* refactor(store): convert singleton to Store.new() factory

Problem: store.lua used module-level _data singleton, making
project-local stores impossible and creating hidden global state.

Solution: introduce Store metatable with all operations as instance
methods. M.new(path) constructs an instance; M.resolve_path()
searches upward for .pending.json and falls back to
config.get().data_path. Singleton module API is removed.

* refactor(diff): accept store instance as parameter

Problem: diff.apply called store singleton methods directly, coupling
it to global state and preventing use with project-local stores.

Solution: change signature to apply(lines, s, hidden_ids?) where s is
a pending.Store instance. All store operations now go through s.

* refactor(buffer): add set_store/store accessors, drop singleton dep

Problem: buffer.lua imported store directly and called singleton
methods, preventing it from working with per-project store instances.

Solution: add module-level _store, M.set_store(s), and M.store()
accessors. open() and render() use _store instead of the singleton.
init.lua will call buffer.set_store(s) before buffer.open().

* refactor(complete,health,sync,plugin): update callers to store instance API

Problem: complete.lua, health.lua, sync/gcal.lua, and plugin/pending.lua
all called singleton store methods directly.

Solution: complete.lua uses buffer.store() for category lookups;
health.lua uses store.new(store.resolve_path()) and reports the
resolved path; gcal.lua calls require('pending').store() for task
access; plugin tab-completion creates ephemeral store instances via
store.new(store.resolve_path()). Add 'init' to the subcommands list.

* feat(init): thread Store instance through init, add :Pending init

Problem: init.lua called singleton store methods throughout, and there
was no way to create a project-local .pending.json file.

Solution: add module-level _store and private get_store() that
lazy-constructs via store.new(store.resolve_path()). Add public
M.store() accessor used by specs and sync backends. M.open() calls
buffer.set_store(get_store()) before buffer.open(). All store
callsites converted to get_store():method(). goto_file() and
add_here() derive the data directory from get_store().path.

Add M.init() which creates .pending.json in cwd and dispatches from
M.command() as ':Pending init'.

* test: update all specs for Store instance API

Problem: every spec used the old singleton API (store.unload(),
store.load(), store.add(), etc.) and diff.apply(lines, hidden).

Solution: lower-level specs (store, diff, views, complete, file) use
s = store.new(path); s:load() directly. Higher-level specs (archive,
edit, filter, status, sync) reset package.loaded['pending'] in
before_each and use pending.store() to access the live instance.
diff.apply calls updated to diff.apply(lines, s, hidden_ids).

* docs(pending): document :Pending init and store resolution

Add *pending-store-resolution* section explaining upward .pending.json
discovery and fallback to the global data_path. Document :Pending init
under COMMANDS. Add a cross-reference from the data_path config field.

* ci: format

* ci: remove unused variable

lua/pending/complete.lua

@@ -15,10 +15,13 @@ end
 
 ---@return string[]
 local function get_categories()
-  local store = require('pending.store')
+  local s = require('pending.buffer').store()
+  if not s then
+    return {}
+  end
   local seen = {}
   local result = {}
-  for _, task in ipairs(store.active_tasks()) do
+  for _, task in ipairs(s:active_tasks()) do
     local cat = task.category
     if cat and not seen[cat] then
       seen[cat] = true