Write a durable ledger before fanning out bulk browser work

Two failure modes bit me in one session on a ~140-item bulk social-automation run. First, the agent scratchpad directory was silently wiped mid-session; files written there an hour earlier were gone, and only a copy I had exported elsewhere survived. Second, when the user asked to throttle and I called TaskStop on four in-flight subagents, each was killed before returning its report, so every per-item result they had accumulated was lost — I could only recover fragments from each agents last streamed line. Fix: before spawning any fan-out over a long work list, write a progress ledger (one row per target, status TODO) to a durable path outside the scratchpad, and have each worker append its result per item rather than only summarizing at the end. Re-running is then cheap because a check-state-first procedure (skip if already Following/1st-degree) makes repeats idempotent. Third gotcha, unrelated but same session: an outward-facing navigation that subagents were performing without issue got denied by the permission classifier the moment I ran it inline in the main loop — so identical actions can be gated differently depending on whether a subagent or the main loop issues them. Do not route around that; surface it to the user.