Detect early JIT compilation issues in CI/CD

Revert batched icache invalidation optimization due to intermittent Linux
kernel boot failures on x86_64 CI runners. The same commit shows both
passing and failing runs, indicating a timing-dependent race condition or
memory ordering issue exposed by the optimization.

Evidence:
- Run 19212176228 (commit 49cd220): SUCCESS
- Run 19212176313 (commit 49cd220): FAILURE (boot test hangs)
- Local macOS ARM64: SUCCESS

The optimization provided 5.6x ARM64 JIT compilation speedup but triggers
intermittent hangs in system emulation boot tests on x86_64 Linux. Further
investigation needed to identify root cause before re-enabling.

This reverts commit a252f41.

Author: jserv

src/jit.c

@@ -397,13 +397,7 @@ static void emit_bytes(struct jit_state *state, void *data, uint32_t len)
     pthread_jit_write_protect_np(false);
 #endif
     memcpy(state->buf + state->offset, data, len);
-    /* Defer icache invalidation to end of block compilation for performance.
-     * Rationale: sys_icache_invalidate() on ARM64 is expensive (~50-100
-     * cycles). Calling it per-instruction during compilation wastes ~80% of JIT
-     * time. Single invalidation after block completion is sufficient for
-     * correctness. Jump patching (update_branch_imm, resolve_jumps) still
-     * invalidates locally.
-     */
+    sys_icache_invalidate(state->buf + state->offset, len);
 #if defined(__APPLE__) && defined(__aarch64__)
     pthread_jit_write_protect_np(true);
 #endif
@@ -2448,21 +2442,6 @@ void jit_translate(riscv_t *rv, block_t *block)
         goto restart;
     }
     resolve_jumps(state);
-
-    /* Batched instruction cache invalidation for entire compiled block.
-     * Performance optimization: Instead of invalidating after each instruction
-     * emit, we invalidate the entire block once after compilation completes.
-     * Impact: ~80% reduction in JIT compilation time on ARM64 (50-100 cycles
-     * per instruction avoided). On x86_64, sys_icache_invalidate is a no-op
-     * (coherent I-cache), so this only improves code clarity without
-     * performance impact. Correctness: Jump patching (update_branch_imm,
-     * resolve_jumps) already invalidates modified locations, so self-modifying
-     * code is handled correctly.
-     */
-    size_t block_size = state->offset - block->offset;
-    if (block_size > 0)
-        sys_icache_invalidate(state->buf + block->offset, block_size);
-
     block->hot = true;
     rv_log_debug(
         "JIT: Translation completed for block pc=0x%08x, offset=%u, size=%u",