fix(repl-jit): fn redefinition rebinds instead of being silently ignored

[Ch4s3]

Problem

At the REPL, redefining a function was silently ignored on both JIT backends:

march(1)> fn f(x) do x + 1 end
march(2)> fn f(x) do x + 100 end
march(3)> f(1)
= 2            <- should be 101

The second definition printed val f = <fn> but had no effect. Interpreter mode
(MARCH_REPL_INTERP=1) already rebound correctly (= 101), so the JIT paths
diverged from the language's Elixir-style rebinding semantics.

run_decl's is_fn_decl path early-returned whenever bind_name was already in
compiled_fns — a guard added for :reset scroll-replay, where the scroll system
resends prior cells verbatim and recompiling is pure waste. The guard could not
tell a verbatim resend from a genuine redefinition.

Approach

Redefinition rebinds, matching interpreter mode. The mechanism the JIT already uses
makes this cheap: cross-fragment calls to a REPL-defined fn resolve through its
persistent closure slot (march_repl_get + indirect call at closure+16 —
verified in the emitted fragment IR), not through the LLVM symbol, so rebinding the
slot redirects every later call.

• Fingerprint each fn declaration (Digest of the marshaled AST), recorded only
  after the fragment compiles and loads — the same success-only discipline as
  mark_compiled_fns, so a failed compile leaves state intact for a clean retry.
  Identical resend → replay fast path preserved; changed body → recompile and rebind.
• Unique symbol for a redefinition (f$redef$<n>), because the earlier fragment
  already defined f and ORC's single JITDylib hard-errors on duplicates (clang+dlopen
  would merely shadow ambiguously in the flat namespace). A capture-avoiding
  rename_top_fn_refs over the post-defun TIR makes self-recursion and first-class
  self-references follow the rename; locally bound shadows are left alone.
• Filter the old slot out of prev_slots so the fragment emits no colliding slot
  loader under ORC. This is what makes the fix independent of the unmerged ORC
  loader-linkage work on claude/jit-repl-interpreted-perf-68373c; the two compose.
• Names compiled outside run_decl (stdlib prelude, :load) keep the historical
  skip — they are called as direct externs, not through a slot, so a redefinition
  could never take effect there anyway.

Tests

• test/test_codegen.ml — in-process (clang): original body, redefinition
  (101 not 2), identical replay compiles no new fragment and keeps the binding,
  self-recursive redefinition reaches the new body, arity-changing redefinition.
  Verified red pre-fix (Received: "2").
• test/test_jit.ml — new repl_session group: subprocess end-to-end sessions of the
  real binary for clang, ORC (MARCH_JIT_BACKEND=orc, skips when libLLVM is absent),
  and interpreter mode. Subprocess because a backend SIGSEGV would otherwise take down
  the runner, and because the backend env vars are read once at module init. Both JIT
  sessions verified red pre-fix; interpreter green before and after.
• New Repl_jit.fragment_count exposes the replay-skip assertion to tests.

Full suite green: compiler 932, eval 272, codegen 588, stdlib 868 (incl. Slow),
stdlib_march 61, jit 4, snapshots 33. scripts/check-docs.sh passes.

Pre-existing bugs found, not fixed here

Both reproduced on unmodified HEAD (68dff6e) with a fresh HOME, and filed
separately:

1. A fn calling a previously REPL-defined fn (fn g(x) do f(x) end) fails on both
   backends — the slot loader define @f() collides with the declare @f extern in
   the same fragment. This is the "Follow-up noticed" item on the ORC-fix branch.
2. A fn whose body contains any lambda fails its helpers-fragment dlopen
   (symbol not found in flat namespace '_f') and loses the binding.

Because of (1), later REPL functions cannot reference f at all today, so there are
no stale direct-call sites for a redefinition to miss. If that limitation is lifted,
direct extern calls inside other functions would pin the version compiled at their
definition time and should be routed through the slot instead — noted in the spec.

Details in specs/progress/2026-08-24-repl-jit-fn-redefinition-silently-ignored.md.