Feedback loop: 👍/👎 on remote matches

[FARADTechnologies]

Context

The remote track currently sends LLM-scored matches, but nothing captures whether the
user actually liked them. Without that signal the matching quality can never improve
beyond the static prompt.

Proposal

• Send high-score matches as individual messages carrying inline 👍 / 👎 buttons
• Handle the callback in the existing Telegram update loop (jobbot/onboarding.py
  already consumes getUpdates; coordinate so both consumers share one offset)
• Persist verdicts to data/remote_feedback.json
• Feed recent 👎 patterns back into the judge prompt (jobbot/judge.py::PROFILE_SPEC)

Notes

Keep the batched list for the long tail; only top matches get their own message so the
chat does not turn into spam.

Acceptance criteria

• Tapping 👍/👎 records the verdict and acknowledges in-chat
• Feedback file survives CI runs (committed like the seen-stores)
• Judge prompt includes a compact summary of disliked patterns