D280 - The first six steps of check could hide the other ten

decided · 2026-08-25 · found by one guard failing and taking the gate with it

D192 established the shape: a tree with six problems should report six, not one per run. must implements it - each step runs in a tested context, failures accumulate, the exit status still reflects them.

Six steps were not going through it. provenance, decisions, prose, generated_numbers, symbols_audit and tables are called bare, so under set -e the first one to fail ends the script - and cargo fmt, clippy, check and the entire test suite never run.

That is worse than a missing report. It is a gate that exits having tested almost nothing, after printing one line about one file, and the run reads as though the tree was examined. A session ended with a watchpoint implementation whose tests had not been run, believing the gate had looked at it, because a symbol database in a directory the implementation never touches was a megabyte too large.

also wraps a bare step the way must wraps a command, for the steps that print their own heading. Same accumulation, same exit status, no stopping.

And that alone was not enough, which is worth recording because the half-fix looked like a whole one. Every one of those six steps reports failure with die, and die calls exit

• so wrapping the call changes nothing, because there is no return value to catch. The gate stopped at exactly the same place with also in front of it.

So bad sits beside die: same red line, return 1 instead of exit 1. die stays for the cases where stopping is right - an unknown verb, a missing binary, a run with no such title - because those are setup problems where continuing would produce noise rather than findings. A gate step is the opposite: continuing is the entire point.