fix(types): check no longer stack-overflows on functions returning T[k]

[JonathanTurnock]

Problem

luabox check aborted the whole run with a stack overflow on ordinary Lua. Seen on Penlight (pl/data.lua) and lazy.nvim (lua/lazy/core/plugin.lua, lua/lazy/manage/semver.lua). Minimal repro:

local S = {}
function S.f(key)
  return S[key]
end

S[key] widens to every value of S, and f is one of them, so f's inferred returns contain f itself. reify_func had no cycle guard (unlike reify_shape), so turning f into a Ty recursed forever.

Fix

• d43e2bf: added a cycle guard to reify_func.
• fd0703e: the guard alone left the acyclic case factorial. With n sibling functions that each return S[k], check took 5.2 s at n=8. Replaced the stack with a depth counter (MAX_NESTED_FUNC_RETURNS = 1): a function's returns are spelled out only at the outermost level.

siblings	before	after
8	5.2 s	0.08 s
30	—	0.12 s
100	—	2.7 s (unavoidable size of 100 unions of 100 members each)

Trade-off

A returned closure's own returns reify as unknown. A factory now hovers as fun(): fun() instead of fun(): fun(): integer. Same-file call checking walks the inference graph and is unaffected. A cross-file consumer calling M.make()() now sees unknown for the inner result. The fd0703e commit message says call checking is unaffected; that holds only for same-file calls. The cap is one constant if we'd rather pay the cost.

Verification

• 3 new tests in luabox-types: the self-reference, 12 siblings (would never finish under factorial cost), and the pinned depth-cap display.
• cargo test --workspace: 3078 passed, 0 failed. clippy -D warnings and fmt --check are clean.
• Real code with the release build: Penlight finishes (strict: 492 errors, almost all from strict mode's handling of untyped values; non-strict: 0 errors, 58 warnings). lazy.nvim finishes with 0 errors and 355 warnings. All 7 examples/ still check clean.

🤖 Generated with Claude Code