A metatable __index function overrides a declared ---@type on main (Neovim's vim.env is "always falsy") #1319

On current main (56267ead, emmylua_check 0.30.0), reading a field of a table declared
---@type table<string, string?> produces the type inferred from its metatable's __index function instead
of the declared value type. That inference can also be wrong. Together they make every
if vim.env.X then in Neovim code an unnecessary-if ("always falsy"), because Neovim's runtime declares
vim.env exactly this way (runtime/lua/vim/_core/options.lua). 0.25.1 reports nothing.

def.lua:

G = {}

--- @type table<string, string?>
G.envA = setmetatable({}, {
  __index = function(_, k)
    local v = os.getenv(k)
    if v == nil then return nil end
    return v
  end,
})

--- @type table<string, string?>
G.envD = setmetatable({}, { __index = function(_, k) return 'always' end })

use.lua:

if G.envA.X then print 'A' end  -- main: Impossible `if` statement: this condition is always falsy
if G.envD.X then print 'D' end  -- main: Unnecessary `if` statement: this condition is always truthy

Expected: both are string? (the declared type), so neither if is reported. That is what 0.25.1 does.

Two separate problems seem to be involved:

1. The declared type loses. envD is declared string?, but reads as string (always truthy). An
   explicit ---@type on the table should take precedence over what the metatable's __index returns.
2. The __index return is inferred as nil. envA's function returns nil on one path and v
   (narrowed to string) on the other, which should be string?. It comes out as nil. With
   return os.getenv(k) the same table reads as string? and nothing is reported.

Controls: setmetatable({}, {}) with no __index, and a plain {}, both keep the declared string?.

Not bisected: 0.25.1 → main crosses the semantic_db rewrite.

Written with help from Claude Code.