> It is quite random.  It happened in a few hours.  It happened after a
 > few days.  I don't remember if it happened after 4 days.  I'll keep
 > collecting more data.

I didn't verify this but considering the recent thread on
`sgml-mode-point-entered' I believe something similar might occur here.
Look at the following two functions:

(defun table--point-entered-cell-function (&optional old-point new-point)
   "Point has entered a cell.
Refresh the menu bar."
   (unless table-cell-entered-state
     (setq table-cell-entered-state t)
     (setq table-mode-indicator t)
     (force-mode-line-update)
     (table--warn-incompatibility)
     (run-hooks 'table-point-entered-cell-hook)))

(defun table--point-left-cell-function (&optional old-point new-point)
   "Point has left a cell.
Refresh the menu bar."
   (when table-cell-entered-state
     (setq table-cell-entered-state nil)
     (setq table-mode-indicator nil)
     (force-mode-line-update)
     (run-hooks 'table-point-left-cell-hook)))

Both do a `force-mode-line-update' which might recursively toggle
`table-cell-entered-state' for preceding cells on the same line.  In
that case the bug should get caught by Yamamoto's fix.  I would
nevertheless apply the attached patch.