claude-tokens watch never re-arms after compaction, so a Claude Code session goes silent for the rest of its life

advance() in bin/claude-tokens only ever adds to w["armed"] — nothing removes a threshold once it has fired, and nothing detects compaction. So when a Claude Code session compacts, usage drops back to (say) 12% and climbs the whole window again, but 50/75/90/95 are all still armed and no further crossing is ever reported. The watch keeps running and keeps printing nothing.

Pi does not have this: token-watch.ts handles session_compact and calls TokenWatch.reset(), so its milestones re-arm and the second climb reports normally. The two implementations disagree about what a milestone means across a compaction.

The Python watcher is an external process polling telemetry snapshots, so it has no compaction event to hook. It would have to infer one — a reading whose percentage drops materially below the previous reading is a compaction (or a session restart, which wants the same treatment), and should re-arm every threshold and re-emit a start line.

Related: #32 adds reminder text to breakpoints. That raises the cost of this bug — a missed re-fire stops being a missed statistic and becomes a missed instruction, and post-compaction is exactly when the agent has lost the context that would have reminded it anyway. #32 is deliberately not fixing this; it is its own defect with its own semantics.

Start in bin/claude-tokens at advance(), then compare Pi's token-watch.ts reset behavior and the described telemetry flow. Done means a material percentage drop re-arms every threshold and emits a new start line, with the watcher reporting crossings again on the next climb.