Add TSan stress suite for per-hotspot race detection Five new scenarios under test/irmin-pack/test_tsan_stress/ target mutable state hotspots that the existing multicore + QCheck-STM tests do not exercise across domains: - stress_mem_cache: races the global Hashtbl.t cache captured by Irmin_mem.Read_only.v (irmin_mem.ml:44) and the KMap mutable in the Read_only instance. - stress_watch: races the listen_dir_hook ref in watch.ml:28-29. - stress_ao_buf: races the unguarded Buffer.t in the rw_perm of Append_only_file (append_only_file.ml). - stress_dict: races the two unguarded Hashtbl.t caches plus last_refill_offset in dict.ml. - stress_fs_pool: races the shared Eio_pool instances in irmin_fs_unix.ml (mkdir_pool, openfile_pool). mem and watch produce clean TSan data-race warnings pointing at the exact hotspot; ao, dict and fs cause memory corruption fast enough that TSan fires via SEGV before it can write a detailed report — the SEGV itself is the race signal. The @tsan-stress dune alias runs each scenario in its own process via `|| true` so that a race-induced crash in one does not suppress the others. The tsan.yml workflow counts both "WARNING: ThreadSanitizer" and "ERROR: ThreadSanitizer" as findings, and includes the @tsan-stress build-dir reports in the uploaded artifact. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>