2026-08-19 (night) - Making the loop canonical

The question was whether a fresh agent on a fresh machine could be handed one command and get on with it. Nearly - and the gaps were worth finding. D079-D081; 273 tests.

• ./orbistoun.sh run <title> is now the single command: resolve a title id, rebuild, refresh names only if stale, run, report, and say whether it went further than last time.
• ./orbistoun.sh doctor says whether a machine can do the work, with install commands. run and names invoke a silent version of it first.
• CLAUDE.md opens with three commands instead of pointing at eighty decisions.

Surprises

• The loop had no objective function, and I had not noticed. It said what a guest wanted and nothing about whether a change helped. That is the difference between iteration and repetition, and no amount of documentation would have fixed it - it needed a number. The faulting instruction pointer is that number.
• run was unusable as a verb. It was a raw passthrough, so debugging a title meant ./orbistoun.sh run run titles/PPSA04263-app0/eboot.bin. Nobody had typed it because nobody had tried to use it the way a person would.
• A grep-based count in doctor was silently double-counting, reporting 528 names for a 264-name database, because derivations sit at the same indentation as names. Dropped rather than fixed: a number that is quietly wrong is worse than no number.
• Always-rebuild would have been the wrong default. Ten seconds per debug run on an unchanged tree is enough that people route around it, and then they debug against stale names - worse than either option the staleness check was choosing between.

Outstanding

Implementing a function is still work, not a command. The loop identifies the wall, measures whether it moved, and keeps the record; it does not write the implementation, and nothing here should pretend it will.