kraken: Fixing undefined behaviour and crash of A*

[AurelienLP]

This PR MUST be merged AFTER #2843.

Only the commit "Fixing A* init by checking if starting edge is found" needs to be reviewed since the first one is the un-revert of the A*.

During A* init, we initialize costs vector for the starting edge. However, if this edge does not exist, its index is either undefined or -1 (max of std::size_t) creating crash or undefined behavior.

We fixed that checking if starting.edge is found. We also access cost vector with "at" and not the operator "[]" which checks the index and the bound of the vector, throwing an exception instead of crashing.

JIRA: https://jira.kisio.org/browse/NAVITIAII-2761

[pbougue]

Just reviewed last commit.

We may as well just keep this PR and not #2843 to avoid remerging unfixed A*

As mentioned orally :

1. It can be worth running artemis + artemis_idfm on that (let you decide)
2. Just curious of impact on perf if you know it (probably not worth rerunning bench just for that)

[AurelienLP]

1. Currently building the PR to run artemis + artemis_idfm
2. It should be negligible since we are not in a loop

[codecov-io]

Codecov Report

Merging #2846 into dev will decrease coverage by 0.8%.
The diff coverage is 94.4%.

@@           Coverage Diff           @@
##             dev   #2846     +/-   ##
=======================================
- Coverage   87.6%   86.8%   -0.9%     
=======================================
  Files        438     446      +8     
  Lines      66672   60121   -6551     
=======================================
- Hits       58433   52192   -6241     
+ Misses      8239    7929    -310

[nonifier]

This is a good fix, but it doesn't prevent us from doing the same mistake ever again.....
I would be in favour of raising an exception in ProjectionData::operator[] if found isn't set.
Or you could change the way ProjectionData::vertices is implemented and make sure it's always in a state that will not bring you to Undefined Behavior Land.

[AurelienLP]

Just updated my PR to throw an exception in ProjectionData::operator[] when found is false ! :D

[pbougue]

Not sure: Is it possible to add a test that would trigger the bug before merge?
I think of 3 "situations":

1. before any fix : doesn't crash (or just using sanitizer) : return wrong result?
2. using .at() crashes
3. after all fix : doesn't crash, return no-result as expected?

It may be nice to have the "functional case" that triggered the bug pinned in a test.

[AurelienLP]

I am trying to create an artemis test for this PR. However, I couldn't reproduce the crash without the fix. We can only create a test that returns "no solution found for this journey" when the starting edge is not found (and check that no crash happens). This is what I am trying to do now :)