Issues found by claude

Verified on Julia 1.10, 1.11, and 1.12 where a runtime claim is made.

Bug

• produce of a non-Int literal whose result is used across a later produce crashes TapedTask construction. src/refelim.jl:234,250 (emitted at src/transformation.jl:812).

  function model()
      s = produce("hi")   # constant literal as produce arg
      t = produce(s)      # result `s` live across a later produce -> ref is kept
      return (s, t)
  end
  TapedTask(nothing, model)
  # ERROR: Unexpected value argument to set_ref_at!: hi
  

  eliminate_refs only handles set_ref_at! value-args that are ID, GlobalRef, or
  Argument; a literal that survives constant-propagation into a kept cross-suspension ref
  hits error(...). The Int case is masked because Core.Const folds the literal into all
  use sites, leaving the produce-result SSA unused (no ref kept). Fix: handle a non-ID
  constant value-arg like the existing GlobalRef/Argument arm (store it as-is for a kept
  set / substitute it for a dropped get) instead of erroring, in both branches.

Docs / CI

• Doc build fails: BasicBlockCode.remake_enter docstring not indexed. docs/src/internals.md.
  makedocs(; modules=[Libtask]) enforces checkdocs=:all; remake_enter has a docstring
  but no @docs entry → [:missing_docs], deterministically on all versions, so Documentation
  CI is red. Fix: add Libtask.BasicBlockCode.remake_enter to the @docs block.
  (Already fixed on the support-trycatch-194 branch; listed for completeness.)

• replace_captures docstring references nonexistent build_rrule (a Mooncake
  leftover) and has a typo "intepreter". src/utils.jl:10. It's a documented internal, so the
  stale text renders in the docs. Fix: reword to Libtask's actual use (reusing a compiled
  MistyClosure's captures on copy/fresh_copy to avoid recompilation).

• optimise_ir! docstring signature is wrong. src/utils.jl:35. Shows
  optimise_ir!(ir, show_ir=false) (positional) but the real signature is
  optimise_ir!(ir; show_ir=false, do_inline=true) (keyword-only; do_inline undocumented).
  Documented internal → the rendered call form throws MethodError. Fix: correct the
  signature and document do_inline.

Robustness

• consume leaves taped_globals in the calling task's TLS. src/copyable_task.jl:520.
  consume sets task_local_storage(TASK_VARIABLE_KEY, …) on the caller and never clears it,
  so after any consume(t), get_taped_globals(::Type{T}) called outside a TapedTask stops
  throwing NotInTapedTaskError and returns the stale value — contradicting both docstrings.
  (The existing test passes only because it is ordered first.) Fix: save/clear the key in a
  try/finally around the inner call, or weaken the docstrings to state the key persists.

• get_function(x::Expr) = eval(x) runs arbitrary IR sub-expressions. src/transformation.jl:67,
  from stmt_might_produce's :call branch. Result only feeds a conservative
  isa(f, Union{IntrinsicFunction,Builtin,DataType}) test, so eval is both unnecessary and a
  side-effect/world-age hazard. Fix: for an unresolved Expr callee, conservatively return
  true (might-produce) instead of evaluating it.

• No _find_id_uses!(::Switch) method. src/bbcode.jl:699-724. characterise_used_ids
  would under-report IDs used only in Switch.conds (latent today: Switch is synthesized after
  the single call site and removed before lowering, but replace_ids does handle Switch, so
  the asymmetry reads as a bug). Fix: add a Switch method mirroring IDPhiNode, or comment
  the invariant.

• necessary_ref_ids dedup relies on a non-local invariant. src/refelim.jl:184.
  vcat(length(refs), …) force-includes the resume-block ref; correct only because that ref is
  never touched via get_ref_at/set_ref_at!. Fix: use a set union so dedup is explicit.

Concurrency / contract

• set_taped_globals! over-restricts the new value's type, contradicting the docs.
  src/copyable_task.jl:499,217. The method …(t::TapedTask{T}, ::T) and the parametric field
  taped_globals::Ttaped_globals pin the type to construction time, so
  set_taped_globals!(TapedTask(nothing, f), 5) throws a MethodError. The docstring says "set …
  to anything you like." Fix: either widen the field/signature, or correct the docstring (and
  the wrong signature shown in the set_taped_globals! docstring) to state the type is fixed.

• generate_ir mutates the shared _id_count outside build_callable_lock.
  src/copyable_task.jl:138. It calls seed_id!() and runs the full IR-deriving pipeline
  unlocked — a second derivation path outside the lock the concurrency note (#227) forbids.
  Debug-only, but violates the invariant. Fix: wrap the ID-generating body in
  build_callable_lock.

• GlobalMCCache's internal ReentrantLock is dead and its docstring misattributes
  thread-safety. src/copyable_task.jl:228-243. All cache access is already under
  build_callable_lock; only setindex! takes the inner lock (haskey/getindex don't), so it
  never contends and the comment points at the wrong mechanism. Fix: drop the inner lock (thin
  Dict wrapper) or lock all accessors + correct the docstring to credit build_callable_lock.

Performance (build-time only)

• get_block_id_from_int linear-scans per resume site → ~O(blocks × resume_sites).
  src/refelim.jl:57. Fix: build an int → ID dict once before the loop.

• Redundant copy of an already-fresh slice. src/refelim.jl:321 —
  vcat(typeof(new_refs), copy(ir.argtypes[2:end])); the slice already copies. Fix: drop copy.

Tests

• allocs perf flag is a silent no-op for empty-result cases. src/test_utils.jl:52-64.
  The allocation loop iterates iteration_results, so cases with [] expected results assert
  nothing — yet "no produce", "kwarg tester 1/2", and both "default kwarg tester" are marked
  allocs. Fix: loop 1:(length(results)+1) (also measure the terminating consume), or
  downgrade those cases to none.

• Duplicate test name + redundant case. src/test_utils.jl:217-227. Two cases share the
  name "default kwarg tester"; kwargs=nothing and kwargs=(;) construct the identical task.
  Fix: rename / drop the redundant one.

Trivial

• Dead assignment new_argtypes = copy(ir.argtypes) overwritten on the next line. src/transformation.jl:979 — delete it.
• throw(error(msg)) double-wraps (and is inconsistent with sibling bare error(...) sites). src/transformation.jl:77,692 — use error(msg).
• might_produce opt-in API documented but not exported/public. src/Libtask.jl / docs/src/index.md:31-35. Consider public might_produce, might_produce_if_sig_contains, @might_produce (guarded for Julia ≥ 1.11).
• 1.12.0-prerelease @warn has no maxlog → floods logs in tight TapedTask-construction loops. src/copyable_task.jl:437 — add maxlog=1.
• Typo "hanled" → "handled". src/bbcode.jl:646 (inherited from Mooncake).

This issue bundles runtime, documentation, concurrency, performance, and test concerns across src/refelim.jl, src/transformation.jl, src/copyable_task.jl, src/bbcode.jl, src/test_utils.jl, and docs/src/internals.md. Start by splitting the checklist into focused changes, then read the cited code and run the relevant tests plus the documentation build. Done means each selected item has a targeted fix, regression coverage where applicable, and passing CI.