fix: mmio: don't report "Intel never assessed this CPU" when the MSR is unreadable

When IA32_ARCH_CAPABILITIES (0x10a) can't be read from userspace (no msr
module, or kernel lockdown under Secure Boot), the FBSDP_NO/PSDP_NO/SBDR_SSDP_NO
bits were left at 0 ("explicitly not immune") instead of -1 ("unknown"). For a
recent CPU not in any kernel model list (e.g. Arrow Lake), this wrongly flipped
the MMIO Stale Data verdict into the "out of servicing period, Intel never
assessed this CPU" bucket.

(cherry picked from commit 23ea5427b5)

src/libs/400_hw_check.sh

@@ -987,8 +987,33 @@ check_cpu() {
                     pstatus yellow NO
                 fi
             elif [ $ret = $READ_MSR_RET_KO ]; then
+                # the MSR access faulted: the register is genuinely absent, so the
+                # pre-seeded 0 ("not advertised") values are correct.
                 pstatus yellow NO
             else
+                # RET_ERR (no msr module) or RET_LOCKDOWN (MSR reads restricted):
+                # CPUID told us the MSR exists but we couldn't read it, so its bits
+                # are undetermined, not 0. Leaving them at 0 would falsely claim the
+                # CPU "explicitly indicates not immune".
+                # Reset every arch-cap-derived value to -1 (UNKNOWN) instead.
+                cap_rdcl_no=-1
+                cap_taa_no=-1
+                cap_mds_no=-1
+                cap_ibrs_all=-1
+                cap_rsba=-1
+                cap_l1dflush_no=-1
+                cap_ssb_no=-1
+                cap_pschange_msc_no=-1
+                cap_tsx_ctrl_msr=-1
+                cap_gds_ctrl=-1
+                cap_gds_no=-1
+                cap_rfds_no=-1
+                cap_rfds_clear=-1
+                cap_its_no=-1
+                cap_sbdr_ssdp_no=-1
+                cap_fbsdp_no=-1
+                cap_psdp_no=-1
+                cap_fb_clear=-1
                 pstatus yellow UNKNOWN "$ret_read_msr_msg"
             fi
         fi