Most of today went into marine route planning — the biggest single push by far, north of two thousand lines. The arc there was one long fight with island navigation. First I tried to make routes conform to real channel and buoy layouts, following the IALA-A marks the way a boat actually would. That crashed mid-build — a worker got killed halfway through — so I came back and fixed the crash, then made the buoyage overlay apply everywhere instead of only kicking in on certain legs.
Then came the real direction change. Rather than keep patching the old approach, I rebuilt the router around a different idea entirely: parsing actual point soundings from the charts so the depth data itself opens up navigable channels. That turned out to be the keystone — it made a harbour that had been stubbornly unroutable actually routable. On top of that I taught the router to go around islands the long way when a leg gets stuck, and to thread buoyed channels toward a berth. It ends with a chart-gated acceptance test that checks a route rounds the north of an island and follows the real markers, with the entrance gate derived from the chart, never hardcoded.
Supporting that, I wrote a batch of probes in a dashboard project — tools to jump the live chart to any coordinate, screenshot it, and eyeball soundings and buoyage to confirm the router wasn’t lying to me.
In my ideas repo, a routing sub-task got re-scoped down to just the approach — I killed some over-engineering on the corridor side — and I pushed a verified 58-point loop route out to my navigation stack.
The Cloudy Brain took its usual automated nightly pass.
Where things stand tonight: the router finally rounds islands and reads real soundings, with a chart-gated test holding it honest.