require: cached module result lost after deep require chain (assert:register nil)

[davydog187]

Summary

After requiring a module whose body issues several nested requires and defines many local functions that close over the required value, the local binding ends up with a value that lacks the required module's fields. The result behaves as if package.loaded[modname] was read while still holding the true sentinel that parse_and_execute_module writes before executing the module body.

Discovered while bumping lua to 9344c7cc9835dfa15f1240e6a9de5c9e102b146a (which includes the ? → modname dot-to-slash fix) in tv-labs/platform/sidecar, which ships luassert under priv/lua/.

Repro

From tv-labs/platform/sidecar on dave/sidecar-bump-lua-1.0.0-rc.0 with mix.exs pinned to commit 9344c7cc9835dfa15f1240e6a9de5c9e102b146a:

priv = :code.priv_dir(:sidecar)
paths = "#{priv}/lua/?.lua;#{priv}/lua/?/init.lua"

lua = Lua.new(exclude: [[:require], [:package]])
lua = Lua.set_lua_paths(lua, paths)

Lua.eval!(lua, "require('luassert.assertions'); return 'ok'")

Errors with:

Lua runtime error: Runtime Type Error
  at -no-source-:307:
  attempt to call a nil value (method 'register' on local 'assert')

luassert/assertions.lua:307 is the first assert:register(...) after local assert = require('luassert.assert') at line 9.

Observations

• require('luassert.modifiers') works. Same shape (local assert = require('luassert.assert') then assert:register(...)) but only 19 lines, and the assert:register calls are right after the require — no intermediate local function defs that close over assert.
• require('luassert.assertions') fails at line 307.
• require('luassert.array') fails at line 66.
• require('luassert.spy') fails at line 182.

All three failing modules have the same pattern: local assert = require('luassert.assert') at the top, ~50+ local function definitions in between (whose bodies do assert(...) via the metatable's __call), then the first assert:register(...) is what trips.

At the failing line, assert is non-nil but assert.register is nil. Reference Lua returns the same obj table both times. The value the local ends up holding looks like the true sentinel that parse_and_execute_module writes to package.loaded[modname] before executing the module body — as if the post-execution overwrite at line 724 isn't seen by the outer state.

Related (also fixed in this branch)

• Original blocker: find_module_file (lib/lua/vm/stdlib.ex:754) didn't convert . to / in modname before substituting into the pattern. Fixed in 9344c7c. This new bug is separate.

Sidecar bump tracking

• tv-labs/platform branch: dave/sidecar-bump-lua-1.0.0-rc.0
• Companion sauron bump (no dotted requires, unaffected): https://github.com/tv-labs/platform/pull/8023