how to debug gc crashes correlated to arena chunk size

In some experimental code, I'm observing a deterministic
crash in gc that appears to be related to a growing arena
used for small amounts of temporary data.

Probably I'm doing something wrong, but I'm surprised
that the crash disappears if the arena is either
- rewinded before each task, or
- using a huge size/growth value (> 4 MB or so).

Any ideas for how to approach this?

;; fatal error messages:
;; - no scavenge function for object 0x1 (widetag 0x1)
;; - no size function for object at 0x1000XXXXX0 (widetag 0x1)

;; chunks => crash in gc shortly after:
;;    64k    used =   171,920  waste =  160
;;   128k    used = 2,023,424  waste = 1200
;;   256k    used =   777,824  waste = 1200
;;   512k    used =   688,240  waste = 1200
;;     1M    used = 1,030,576  waste = 1920
;;     2M    used = 2,195,872  waste = 4080
;;     3M    used = 3,218,048  waste = 6000
;;     4M    used = 8,457,792  waste =15920
;;     5M    no crash observed (at least up to 37,328,480 bytes used)
;;    16M    no crash observed (at least up to 37,328,480 bytes used)