[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=97
final_errors=0
chunks=6
model=gpt-5.5
timestamp=2026-06-11T10:24:07Z

Branch Topology

All branches in the BAR type-error cleanup stack. Regenerated deterministically by just bar::fmt-mig-generate. Generated 2026-06-11 10:25:03 UTC.

Leaves — each targets master, mergeable independently:

Branch	Command	Diff vs parent	Units
fmt	stylua	1450 files, +183025 −187662	✅ pass
mig-bracket	bar-lua-codemod bracket-to-dot	342 files, +7653 −7653	✅ pass
mig-rename-aliases	bar-lua-codemod rename-aliases	173 files, +362 −362	✅ pass
mig-detach-bar-modules	bar-lua-codemod detach-bar-modules	178 files, +1694 −1562	✅ pass
mig-i18n	bar-lua-codemod i18n-kikito	20 files, +121 −1198	✅ pass
mig-spring-split	bar-lua-codemod spring-split	782 files, +10019 −10010	✅ pass
mig-integration-tests	<hand curated>	22 files, +140 −90	✅ 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	1480 files, +193205 −197222	897 files, +21353 −20733	✅ pass
fmt-llm-source	1486 files, +193710 −197492	55 files, +575 −340	✅ pass
fmt-llm	1487 files, +193804 −197514	51 files, +117 −45	✅ pass

[github-actions]

Integration Test Results

0 tests   0 ✅  0s ⏱️
0 suites  0 💤
0 files    0 ❌

Results for commit 48d040b.

♻️ 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.