fix(hosts): the handoff report can adopt an earlier live run and name its run_id #1101

Problem

scripts/dsh-superagent-run.sh snapshots every status.json in the store before
the launch (:53-76, mtime + size), then reports the first file whose
mtime/size changed (:82-101, max by mtime at :101). A store is reused per
Ticket (docs/dsh-super-agent.md:165) and a live run rewrites its own
status.json every round (runstore.py:445-455), so a change to an earlier,
still-running entry is indistinguishable from this handoff's new file. The
report then carries found: true with the earlier run's run_id, round and
run_dir, and the page's §6 read-back polls the wrong run.

The page's promise at :166-168 — "found: false until its own run writes
one" — only holds for finished earlier runs, whose file is static. It is false
for a live one, which is the case a shared store hits most often.

Evidence

Measured at live main 09492927, from run-1115-1304/probe-adopt-old-run.sh
(+.txt), three identical runs, tmux, DSH_RUN_HANDOFF_WAIT=3:

Row B is 3/3 identical, exit 0. The reported run_id belongs to a run this
launch did not start, and the read-back that follows polls it.

Slice

The discriminator is already on disk and already written by the runtime:
status.json's started (runstore.py:687, :823), which update_status
never resets (:445-455), so an earlier run keeps its launch time while its
updated and round move.

• take the launch epoch into the snapshot (:55) — time.time() at snapshot
  time, written into the same JSON;
• in the poll (:82-101), accept a changed file only when its started is
  numeric and >= that epoch. A file absent from the snapshot still counts, so
  a fresh run directory and a caller-owned run directory behave exactly as they
  do today;
• tighten the page sentence at :166-168 to: a file whose started predates
  the launch is never reported;
• one row in tests/test_dsh_superagent_run.py after the existing reused-store
  row, in the same gated group, no model;
• one changelog.d/ fragment.

Cost: one field in the snapshot JSON, one comparison in the loop. No new flag,
no extra wait, the wrapper still returns inside WAIT.

Tier: T2 — scripts/** is a T2 path in .github/agent/risk-rules.json;
docs/** and tests/** are lower. Danyang's click, so a PASS at the tip first.

Acceptance

1. Row B's setup: an earlier live run rewriting its status.json every 200 ms,
   this launch writing <store>/<new-run>/status.json at +1.2 s — the report
   names new-run, never old-run, and the read-back at run_dir follows the
   new run.
2. Row B with a child that writes nothing inside the window — the report is
   found: false (the older run is not adopted), same as today's else-branch.
3. Rows A and C unchanged; the existing five rows in
   tests/test_dsh_superagent_run.py stay green, the tmux no-orphan row
   included.

Non-goals

• The two-wrapper waiting.json handshake at :170-173 — unchanged.
• #1096's exit code and waiting state when the child never started. That is
  a liveness decision on the same poll loop; this card is the adoption bug only.
  Both touch :53-101, so whichever lands second rebases.
• The session-name scheme, a --run-id flag, watching or retrying after the
  handoff.

Written-by: AgentDescent Product — found in the 13:04 audit at 09492927, reproducible 3/3.