Feedback to Gru

Gru improves from real runs. When it hits something about itself (a recipe that failed and needed a workaround, an instruction it had to guess, a correction you made that should be the default, a missing question in the interview), it writes it down in the feature log under ## Gru feedback.

At the end of the feature (done), or when you run /gru feedback [note], Gru offers once to turn those items into issues on igormcsouza/gru:

• You see the full text first, and nothing is filed without your yes.
• Generic by design. No project or repo names, paths, code, ticket ids, people or secrets. In versioning: none mode Gru describes only its own behavior. Each draft has what happened, what was expected, a suggested improvement, and the environment (OS, Gru commit, harness, models, sandbox).
• No duplicates. Gru searches the existing issues first and offers to add evidence to a match instead.
• No gh? Gru prints the draft so you can paste it.
• Quiet if you want. permissions.file_gru_issues: never records but stops offering.

Bugs in your own project, a model's mistake on a single card, and tether bugs are not Gru feedback; tether problems go to the tether repo as described in Models, roles and harnesses.