fix(conditions): gate the CLOS-condition tier on a persisted marker (cave#294)Verified

f2767ea's mc probes put a baked-function call (ag-fn-lookup) on the
UNIVERSAL error path. At the host-built stage1, runtime-compiled code
(stdlib loaded at boot, REPL forms) cannot call ANY baked defun — the
call traps and spins — so every condition raise died: handler-case,
condition-system-stubs, bitwise-and-arithmetic, fboundp bootstrap tests
all failed.

Fix (stdlib/mini-clos only; not in COLD_SRCS, zero fixed-point risk):
- New *mc-conditions-live* stdlib defvar; mini-clos SETQs it T at load
  (after mc-conditions-init-builtin — *mc-print-hook* precedent).
- All six f2767ea mc sites (%cond-mc-inst-p, %mc-accessor-fallback,
  %define-condition-runtime, make-condition, both printer arms) check
  the flag — a same-tier defvar read, the only probe safe in every
  tier — BEFORE any ag-fn-lookup. stage1: mc never loads, flag NIL,
  cons-tier conditions (pre-f2767ea behavior). stage2/funcl: flag T,
  instance tier as before.
- %condition-image-tables/-restore-tables gain slot 3 carrying the
  flag across save-image (the cave#291 channel — bake-time values do
  not survive restore, cave#244); old 3-slot images read NIL and the
  mc arms come up safely off. save-image round-trip verified: restored
  image still builds instances (typep T).
- %fn-lookup-hit-p belt kept at the sites (a degenerate lookup result
  can never feed NIL into a numeric compare again).
- fboundp: pure-stdlib special-operator-p arm reordered first and the
  ag-macro-known-p (baked) arm handler-case-wrapped — the same
  stdlib-calls-baked-fn class made (fboundp t) exit 1 in stage1 a.outs.

Verified: stage1 handler-case → BAD / nested → b; 4 of the 5 cave#294
tests pass; native tier unregressed (typep/make-condition T,
handler-case catches, defclass smoke); fasl oracle still 4/4;
make build green.

The residual save-image-restart-evaluates failure is OUT of this
family: stage1 (save-image …) core-dumps at f2767ea~1 too (gdb:
AG-SAVE-IMAGE-ACTION) — an older stage1 save-path breakage, filed
separately, as is the root "stage1 runtime code cannot call baked
defuns" class.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>