daily-fix: dispatch lease races (4 dup spawns + cap 6>5)

Workflow gap

Five same-day incidents show the proposed-infra dispatch path re-checks neither the lease nor the cap at spawn time:

• Duplicate /issue sessions spawned against live owners: #1771 (59 s after the owner's registration), #1997 (26 s after the owner's progress note), #1988 (9 s after the watcher dispatched the owner), #1992 (29 s after owner activity). All four were caught by the Step 0 single-orchestrator guard and exited cleanly — the waste is four spawns + boot cost, and the 9 s gap on #1988 suggests two dispatch lanes firing on the same ripe task in one tick.
• Cap overshoot: after dispatch resumed post-daemon-outage, "Occupancy is 6 (the sweep had dispatched past the old cap during the window I wasn't watching)" with cap 5 — the per-pass batch dispatch does not re-check occupied < cap before each individual spawn (or raced a cap edit).

verified-at-filing: all five are the recovery miners' probed transcript reads (sessions d38f46fa rows 22–27, 644f41fe rows 22–29, 910f48c8 rows 23–31, 5ca7bd0e rows 19–24, 4966e56e rows 1163/1255). ls tests/test_dispatch_lease.py at compose time → the lease test file exists; the sweep path's coverage of it is the question.

Proposed change

In scripts/autonomous_session_watch.py's dispatch loop(s): re-check the dispatch lease immediately before POSTing each spawn (not once per pass), skip candidates whose issue has a session registration younger than the sweep interval, and re-check occupied < cap before each individual spawn inside a batch. Extend tests/test_dispatch_lease.py with the two race shapes (fresh-registration skip; per-spawn cap re-check).

Provenance

• fingerprint: 8cd2ba879258

• workflow_fix_target: scripts/autonomous_session_watch.py

• origin: /daily 2026-08-04 recovery sweep — miner 5 P14, miner 4 P23, miner 7 P12/P14.