fix(jit): a REPL fn calling a prior REPL binding no longer emits invalid IR (both backends)

[Ch4s3]

Symptom

march(1)> fn f(x) do x + 1 end
val f = <fn>
march(2)> fn g(x) do f(x) end
jit error: clang failed (...): invalid redefinition of function 'f'
march(3)> g(41)
I cannot find `g`.

Same failure with a prior let-bound lambda as the callee, on both the clang and ORC backends. Interpreter mode (MARCH_REPL_INTERP=1) was fine. Filed as the "Follow-up noticed, NOT fixed here" section of specs/progress/2026-08-24-orc-repl-segfault-fn-def-after-let-lambda.md.

Root cause

Fn-defining REPL fragments expose prior bindings via module-level slot loaders (Llvm_repl.emit_slot_loader_fns emits define ptr @f()), unlike expression fragments, which bridge slots into var_slot allocas. The EApp emission path knew about neither top_fns nor var_slot entries for f, so the call fell into the unknown-function fallback, which emitted declare ptr @f(i64) for the very symbol the loader defines in the same module — invalid IR, whole fragment (including g) lost.

Note the collision partner is that fallback declare, not partition_fns/extern_fns — f is absent from the second fragment's TIR entirely. Even without the collision, a direct declare+call would be wrong twice over: the erased call-site signature (ptr @f(i64)) mismatches the real fragment-0 define (i64 @f(i64)), and a direct extern call pins the callee version compiled at g's definition time instead of following the slot (the semantics 2026-08-24-repl-jit-fn-redefinition-silently-ignored.md says to prefer).

Fix

Route calls to slot-loader names through the loader + closure dispatch:

• Llvm_ctx.repl_slot_fns — new ctx table; emit_slot_loader_fns records each bare name it emits a loader for. Only REPL fn-fragment emission ever populates it, so non-REPL compilation is untouched.
• emit_atom: new AVar arm — a reference to a slot-loader name emits call <ty> @<name>(), materialising the slot's current value (the closure, for fn/lambda bindings). Placed above the runtime-prefix/builtin first-class arms so a REPL binding shadows a same-named builtin, exactly as the expression-fragment var_slot bridge does.
• EApp: new arm mirroring the existing var_slot→ECallPtr redirect.
• ECallPtr no-var-slot catch-all: excludes slot-loader names so they reach the generic closure-dispatch arm.

No declare is emitted, and the call goes through the closure's $clo_wrap (uniform tagged-ptr ABI — correct for odd/even Ints, Floats, Strings), reading the slot at call time.

Testing

• 4 new subprocess session tests in test/test_jit.ml ({clang,orc} × {fn-calls-prior-fn, fn-calls-let-lambda}), written first and failing 4/4 before the fix; 8/8 green after. The harness was generalised to take a backend (clang sessions don't require libLLVM to run).
• Manual sessions on both backends: chained slot calls (k → g → f), odd/even Int results, String concat through two slot fns, Float chain, recursion + slot call, prior-let capture, let-bound-lambda callee, helper-lambda callee — all correct.
• Full suite: compiler 932, eval 273, stdlib 877, stdlib_march 61, snapshots 33 — green. run_codegen: 587 run with 21 pre-existing failures that are ORC-by-default fallout from the included branch (in-process sessions sharing one LLJIT dylib collide on _Eq$Int.eq in repl_0, before any slot loader exists; all 21 pass with MARCH_JIT_BACKEND=clang). A separate task is under way for that.

Notes

• This branch includes a merge of claude/jit-repl-interpreted-perf-68373c (ORC default + internal-linkage slot loaders + pre-warm); origin's copy of that branch ref was stale, so those commits ride along here.
• Complementary to fix(repl-jit): fn redefinition rebinds instead of being silently ignored #339 (fn redefinition rebinds): that PR makes redefinition rebind the closure slot; this one makes calls read through the slot — together they give interpreter-parity late binding. Whichever merges second should add a session test: fn f → fn g calls f → redefine f → g(x) picks up the new body.