[Types] LLM-driven type-error transform capstone

[keithharvey]

Part of BAR type-error cleanup. Final stage: deterministic transforms + env layer + LLM type-fix pass.

LLM workers apply the categorized fix recipes in SKILL.md — each category maps an emmylua_check error pattern to an idempotent recipe.

Triage run

baseline_errors=56
final_errors=0
chunks=6
model=gpt-5.4
timestamp=2026-07-08T19:09:53Z

Branch Topology

All branches in the BAR type-error cleanup stack. Regenerated deterministically by just bar::fmt-mig-generate. Generated 2026-07-08 19:10:44 UTC.

Leaves — each targets master, mergeable independently:

Branch	Command	Diff vs parent	Units
fmt	stylua	1373 files, +177644 −182508	✅ pass
mig-bracket	bar-lua-codemod bracket-to-dot	333 files, +7578 −7578	✅ pass
mig-rename-aliases	bar-lua-codemod rename-aliases	174 files, +364 −364	✅ pass
mig-detach-bar-modules	bar-lua-codemod detach-bar-modules	178 files, +1688 −1560	✅ pass
mig-integration-tests	<hand curated>	20 files, +134 −81	✅ pass
mig-busted-types	<hand curated>	12 files, +1501 −0	✅ pass

Rollups — composite branches stacking the leaves and (for fmt-llm) the env + LLM layers:

Branch	Diff vs master	Diff vs parent	Units
mig	1398 files, +181877 −185059	509 files, +11219 −9537	✅ pass
fmt-llm-source	1404 files, +182404 −185341	57 files, +581 −336	✅ pass
fmt-llm	1405 files, +182461 −185363	35 files, +74 −39	✅ pass

[github-actions]

Integration Test Results

14 tests  ±0   6 ✅ ±0   3s ⏱️ ±0s
 1 suites ±0   8 💤 ±0 
 1 files   ±0   0 ❌ ±0 

Results for commit 3a80008. ± Comparison against base commit 8869f1e.

♻️ This comment has been updated with latest results.

[keithharvey]

Integration tests are passing, it's a lie because of the engine version. I could zip up the engine and run a custom workflow every time I push this to prove that, but I'm not gonna.