Add version corruption testcase and fix binary corruption

[edwintorok]

For issue 6668 , expected to fail and segfault until the bug is fixed

[edwintorok]

This is the failure that I see (note that the repeated prerr_endline seems important, too few of those and it may hang instead of segfault):

diff --git a/_build/.sandbox/47356ecdc580a3b8a0758a8d78662233/default/test/blackbox-tests/test-cases/versioncorruption.t b/_build/.sandbox/47356ecdc580a3b8a0758a8d78662233/default/test/blackbox-tests/test-cases/versioncorruption.t.corrected
index baa0b6b36..e940baeb3 100644
--- a/_build/.sandbox/47356ecdc580a3b8a0758a8d78662233/default/test/blackbox-tests/test-cases/versioncorruption.t
+++ b/_build/.sandbox/47356ecdc580a3b8a0758a8d78662233/default/test/blackbox-tests/test-cases/versioncorruption.t.corrected
@@ -60,9 +60,12 @@
   0.0.0
   0.0.0
   0.0.0
+  Segmentation fault
   Found placeholder in _build/default/gen_lifecycle.exe:
   - placeholder: Vcs_describe "."
   - evaluates to: "v0.0.1"
   Found placeholder in _build/default/gen_lifecycle.bc:
   - placeholder: Vcs_describe "."
   - evaluates to: "v0.0.1"
+  Segmentation fault
+  [139]

[edwintorok]

I've updated the testcase slightly, so it now also prints a hex diff of original vs promoted binary, to catch corruption even in cases where it doesn't lead to segfault, and with the buggy dune it'd fail like this, the diff would be much bigger as can be seen from the diff header lines (the full diff is not printed because it'd be too big).

I was not able to reproduce the corruption just by using 'dune install', looks like promotion to source dir is required to reproduce it (or perhaps it is just scheduling dependent):

File "test/blackbox-tests/test-cases/versioncorruption.t", line 1, characters 0-0:
diff --git a/_build/.sandbox/9f7765d30bc1cebc3a267aceb684acb3/default/test/blackbox-tests/test-cases/versioncorruption.t b/_build/.sandbox/9f7765d30bc1cebc3a267aceb684acb3/default/test/blackbox-tests/test-cases/versioncorruption.t.corrected
index 71077ddc7..8908e13b9 100644
--- a/_build/.sandbox/9f7765d30bc1cebc3a267aceb684acb3/default/test/blackbox-tests/test-cases/versioncorruption.t
+++ b/_build/.sandbox/9f7765d30bc1cebc3a267aceb684acb3/default/test/blackbox-tests/test-cases/versioncorruption.t.corrected
@@ -79,10 +79,8 @@
   > EOF

   $ rm -f gen_lifecycle.bc gen_lifecycle.exe && dune clean && dune build && ./gen_lifecycle.exe
-  0.0.1
-  0.0.1
-  0.0.1
-  0.0.1
+  Segmentation fault
+  [139]
   $ od -t x1 _build/default/gen_lifecycle.exe >gen_lifecycle.old.txt

   $ dune install -j16 --prefix=./_install 2>/dev/null
@@ -107,13 +105,11 @@
   Found placeholder in _build/default/gen_lifecycle.bc:
   - placeholder: Vcs_describe "."
   - evaluates to: "v0.0.1"
-  0.0.1
-  0.0.1
-  0.0.1
-  0.0.1
+  Segmentation fault
+  [139]

   $ od -t x1 ./gen_lifecycle.exe >gen_lifecycle.new.txt
   $ diff -wu --label old gen_lifecycle.old.txt --label new gen_lifecycle.new.txt | head -n3
   --- old
   +++ new
-  @@ -47632,10 +47632,9 @@
+  @@ -47632,606 +47632,656 @@

[edwintorok]

Although I suppose the offsets in the diff might change with compiler versions.... might have to use 'sed' to extract just the part after the comma (the length) ... updated to extract just the 2nd number in the diff, so it shouldn't depend on the offset anymore.

[edwintorok]

I've made the testcase a bit more robust to work with multiple compiler versions/OSes, and all the tests pass here now.
@rgrinberg this is the PR to fix the binary corruption bug I raised here