fix: reapply tracer circular dependency fix without breaking if-block narrowing

[RomanSpector]

Summary

3.18.2 reverted the tracer fix from #3392 ("Fix type inference for and/or idioms and circular dependency in tracer", 46d0faf) because the _compilingAssigns guard broke type narrowing in if-blocks that reassign the variable. This PR reapplies that fix and resolves the regression that caused the revert, restoring the fixes for #2236, #2374 and #2494 (all three are broken again on current master).

Root cause of the regression that triggered the revert

The guard fired when calcNode re-entered an assign that was still being compiled — this happens when the assigned value's type depends on the variable itself (e.g. x = x .. 's'). On the re-entrant path it skipped lookIntoBlock, so nodes[subBlock] was left unset; the if handler then fell back to the branch entry node, dropping the assignment's effect, and the outer walk cached that poisoned narrowing for every position after the block.

The bug was invisible to the test suite: it only manifests when a position inside the branch is compiled before a position after the block — the normal compile order of an editor session, but something single-query-per-file tests never do. That is why CI stayed green while the editor misbehaved:

---@type integer?
local x

if x then
    x = x .. 's'
    print(x) -- hover here first: string
end

print(x) -- then here: was integer? (assignment lost), should be string?

Fix

calcNode now tracks in-flight assigns per coroutine (weak-keyed by coroutine.running(), so an abandoned coroutine can no longer permanently leak the in-flight flag) and counts re-entrant hits. When the outer call finishes compiling and a re-entrant hit happened meanwhile, the tracer drops its nodes/mark caches, so every later lookup recomputes with the final type instead of the poisoned intermediate state. Purges are rare (only self-referential assignments reachable from control-flow handlers) and terminate because vm.compileNode results are cached, so a source can trigger at most one re-entrant compile.

Tests

• New test/type_inference/reassign_narrow.lua with a dual-query harness that queries a position inside the branch before a position after the block, reproducing the editor compile order:
  • on current master these cases degrade to unknown inside the branch and nil after it (the reverted Type is "lost" of variable within For Loop #2374/Can not infer basic assignments in nested loop. #2494 behavior);
  • with only the original 3.18.1 guard they lose the branch assignment (integer? instead of string?) — the regression described in the 3.18.2 changelog.
• Full test suite passes on win32-x64 (bin/lua-language-server.exe test.lua).

Closes #2236, closes #2374, closes #2494

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.