When I load factory settings it crashes too, but not every time. When I run debug .cmd files, then it seems to crash more often:

I managed to reproduce the problem on a different Windows machine, although it does seem to happen at random:

As I wrote, it doesn't happen in 2.90. In 2.83.1 best way to reproduce it for me is to use debug modes. Then it crashes 100% times:

EDIT:
Just tested on 2.83.2 and it's the same.

Here's a different test with the same file and 2.90 3623db7784b3 (on a different Windows 10 machine). Reopening the file quickly will crash Blender, but that could be a different bug altogether:

I managed to redo a crash with

switching view types, although it crashed when switching to the "3D cursor" in my case.

This is the ASAN output:

@Bastien Montagne (mont29), this looks to be an issue in memfile undo/redo, are you aware which change in 2.90x that would have fixed this?

Fwiw, bisect shows rBcda15408582e: Undo: Detect/find proper memchunk for a given ID during undo step writing. "fixes" the issue... basically by avoiding a lot of useless re-reading (so it actually rather hides it).

In any case, it's again those horrible bone pointers in pose data, crash happens when storing an undo step with dirty pose data needing to be rebuilt. That very ugly quickpatch (based on rBcda15408582) fixes it for me:

diff --git a/source/blender/blenloader/intern/writefile.c b/source/blender/blenloader/intern/writefile.c
index e3b4166a4bf..31e80b9d260 100644
--- a/source/blender/blenloader/intern/writefile.c
+++ b/source/blender/blenloader/intern/writefile.c
@@ -151,6 +151,7 @@
 #include "MEM_guardedalloc.h"  // MEM_freeN
 
 #include "BKE_action.h"
+#include "BKE_armature.h"
 #include "BKE_blender_version.h"
 #include "BKE_bpath.h"
 #include "BKE_collection.h"
@@ -4214,9 +4215,14 @@ static bool write_file_handle(Main *mainvar,
           case ID_AC:
             write_action(wd, (bAction *)id_buffer, id);
             break;
-          case ID_OB:
+          case ID_OB: {
+            Object *object = (Object *)id_buffer;
+            if (object->pose != NULL && (object->pose->flag & POSE_RECALC) != 0) {
+              BKE_pose_rebuild(mainvar, object, object->data, true);
+            }
             write_object(wd, (Object *)id_buffer, id);
             break;
+          }
           case ID_MA:
             write_material(wd, (Material *)id_buffer, id);
             break;

But I would rather avoid such horror... We could also maybe just call BKE_pose_clear_pointers() instead ? Not sure how the read code would react to that though...
I wonder if we should not rather just systematically re-build from scratch pose data in readcode? Would be the safest way to deal with those issues imho.

@Brecht Van Lommel (brecht), @Sybren A. Stüvel (sybren), any thoughts on this?

lib_link_pose already has code for rebuilding the pose and looking up the bone pointers again? If that works correctly there should be no invalid bone pointers after file load.

From what I understand it's this code that is crashing, and it's the only code that is accessing the bone pointer in the file writing code.

/* prevent crashes with autosave,
 * when a bone duplicated in editmode has not yet been assigned to its posechannel */
if (chan->bone) {
  /* gets restored on read, for library armatures */
  chan->selectflag = chan->bone->flag & BONE_SELECTED;
}

BKE_pose_clear_pointers would disable this code, then we might as well remove it.

Something like this I guess will work. Preserving selection here seems important, and I don't immediately have a better solution.

Bone *bone = (pose->flag & POSE_RECALC) ? BKE_armature_find_bone_name(arm, pchan->name) : pchan->bone;
if (bone) {
   chan->selectflag = bone->flag & BONE_SELECTED;
}