A short read no longer condemns a healthy node type

Every portal, on startup, looks up the source files belonging to each of its node types and compiles them. On 2026-09-08 one boot of memex.systemorph.com got a short answer from that lookup: it found 1145 code files where the boots before and after it — three of them running the very same image — each found more than 1236.

The four node types whose files were in the missing part were handed to the compiler with nothing to compile, and the compiler said what it honestly saw: types that do not exist, names that are not in scope. Those messages are indistinguishable from a genuine breakage, so the portal concluded its new image had broken four node types and did the right thing with that conclusion — it refused to take traffic, kept the previous pods serving, and stalled the rollout. Nothing broken was ever served. But nothing was actually wrong, and because nothing was wrong there was also nothing to change that would clear the verdict. The pod sat out of rotation until its startup budget of three hours ran out and it was restarted, at which point the same image started cleanly.

The lookup now asks a second question before it accepts an empty answer. Each node type already records how many source files it has, so "I found none" can be checked against "the mesh says there are two". When those disagree, the pass reports that it could not establish the source set at all, and the portal falls back to resolving each type on its own — slower, and correct — instead of turning a short read into a verdict about the code. A node type that genuinely owns no source files still records that fact, and still classifies as content rather than as a broken image, exactly as before.