[aft-alfonso]

Test-only fix for a merge-queue failure.

• subc::tool_provider::route_tests::occipital_task_observation_is_limited_to_bound_session wrote two sessions' task bundles with task_paths + write_task, replaying each session in between. The first replay's detached GC could remove the second session's still-empty directory before its metadata was written (Linux, PR land: route-project-from-stamp #996's merge group: write_task NotFound). It now allocates through create_task_layout and writes with write_task_at, the same fix the consent scratch-task test got in Merge-queue flake batch 2 #1043.

Build server: -D warnings check; the test passed in 5 of 5 loops.

• Test-only: subc_storm_test published the shared storm root's project config with truncate-and-write on every rebind while a sibling route's bind read it; a torn read failed the bind with config_divergence (Linux release storm, PR land: bash-watch-each-line #911's group). It now writes a temp file and renames it into place, as #891: route binds read config off the frame loop and acknowledge on their own lane #984's fixture does.
• Test-only: the macOS thread inventory in idle_root_retention_tests.rs named a thread's start routine only from saved frames, which never include the running function. A probe sampled while executing its own closure body reported just thread_start > _pthread_start (PR Accept worker codemode runs (worker-runs/v1) with the worker's own authority #1014's group, macOS libtest). It now also consults the program counter and, on arm64, the link register; saved frames still win. A hand-built stack test covers all three cases.