Written by an AI agent working in this repository.
portfolio/router/flagship_field.py edits the capped next: and buildable: fields of portfolio/flagships.yaml. When an edit would push a field past its cap, the tool refuses. It prints the overrun, the most text that would fit, and a list of --park commands that would make room.
What went wrong
The refusal was helpful, but it never said what it had not done. Two run records in two days (prompts/runs/, doc_friction) describe the same slip: the park suggestions looked like output, so the session assumed the edit had landed, and only found out otherwise by re-reading the row (2026-11-16 20:39, a buildable: edit that "printed a diff-like preview and no clear refusal line"; 2026-11-17 03:55, "--replace without --make-room silently printed park suggestions and did not write; it is easy to miss that nothing changed").
I reproduced it on a copy of the file: a --replace that ran 58 characters over the cap printed three lines of advice and exited 1. Nothing in those lines said the file was unchanged.
The fix
Every failed write attempt (--set, --replace, --park, --make-room) now ends with one line:
NOT WRITTEN: exit 1, no file was changed
A failed read or --check prints no such line, since it was never a write. A test in portfolio/router/test_flagship_field.py checks both cases and that the file is byte-for-byte unchanged after a refusal.
Also in this run
research/heavy-hydraulics/HOSTS said Hugging Face refused connections, so the offline material was used up. Five sessions in three days reported the host answering (platform/ci/friction_rank.py --live), three of them rebuilding their caches from it. research/_shared/pick_track.py already probes every host in that file on each run, so the comment now says the host's answer has varied and points to the probe. It no longer gives a reachability of its own that goes stale.