We need @Clément Foucault (fclem) help here.

I only add the render section for select, but I did not make any code of the selection itself. No idea why is working in this way.

Is this bug still reproducible in the latest builds?

I tried again by disabling the OpenGL Depth Picking option and I was able to reproduce the bug.

The problem is that gpencil calls gpencil_free_runtime_data(stl) inside GPENCIL_draw_scene() thus releasing some resources in the first call from DRW_draw_select_loop(...).
However, DRW_draw_select_loop(...) calls drw_engines_draw_scene() at least three consecutive times.
I don't know if the solution has to be done in the DRW_manager itself or in the GPencil engine.

I also noticed an efficiency issue.
gpu_select_query_load_id(id) is called several times consecutively for the same id.
This forces the creation and use of more glQueries.
It would be good to investigate this too.

This forces the creation and use of more glQueries.

If I remember correctly, Similar queries are merged together.

However, DRW_draw_select_loop(...) calls drw_engines_draw_scene() at least three consecutive times.

Best way of fixing it would be to introduce a post draw callback in the draw engine struct and free the runtime data here. But that maybe a bit too big of a modification at this time.

Other possible fix is to move all these resources to a per layer storage (DRW_view_layer_engine_data_get) and discard only in cache_init.

Is there any way to recreate the entire runtime_data within GPENCIL_draw_scene(...)?
It would be another possible fix.

I think this can be complex...you have a list of object sorted by z-depth, the layers, the strokes with different types, etc...not sure what really you need, but maybe could be possible if you give me more details of what you want to do.

Have you checked drw_gpencil_select_render(stl, psl); function?

@Antonio Vazquez (antoniov) would be nice to remove that gpencil_free_runtime_data(stl) from GPENCIL_draw_scene(...).
At least for selection.