Send the user's feedback about s0ma to the s0ma team.

Their words so far: __MESSAGE__

Nothing leaves this machine until the user approves the exact feedback data
in the client's own approval form. Work these steps:

1. Identify the feedback: what went wrong or is missing, what they expected,
   and what happened instead. Ask only for what is genuinely missing.
2. Confirm the target first if their words did not name s0ma or kn0wthyself.
   Other tools collect feedback too, so ask whether this report is meant for
   the s0ma team before going further. Words that name s0ma need no such ask.
3. Draft it concisely from evidence you already have: their own account, the
   diagnostic codes they hit, the tool or command involved. Leave unrelated
   source code and unrelated conversation out — never paste a file, a
   transcript, or a credential into the report.
4. Call prepare_feedback(summary=..., category=..., impact=...) with that
   draft. It saves nothing and sends nothing; it only writes the report up
   and cleans sensitive values out of it locally.
5. Show the user the exact feedback data it returned, and say what its
   redaction report removed. Any edit of theirs means a fresh
   prepare_feedback call — the report and its fingerprint always travel
   together, unchanged.
6. Call submit_feedback with that report and fingerprint unchanged. The
   approval form does the asking: never state or assume that the user
   approved, and never read frustration or silence as a yes.
7. Relay the outcome plainly: the receipt id when it sent, or — on refusal,
   on failure, or in a client with no approval form — the saved draft id and
   the one command that finishes it from a terminal: s0ma feedback.

Done when: the user has seen the exact feedback data, decided for themselves,
and holds either a receipt id or a draft id.
