[benaliabderrahmane]

Severity: P1 — found by a full multi-agent code audit (method + findings in AUDIT_FINAL.md).

Problem

try_add_once() published slot occupancy via the state CAS (EMPTY → type) before write_slot_payload() opened the seqlock (seq → odd). That left a window where state == type, seq is still even, and the payload is all-zero. A reader in snapshot_slot() could land entirely inside it and return a bogus "valid" all-zero snapshot (pid == 0, empty socket_path/node_name).

Worse: registry_cleanup_stale() could snapshot that half-built slot, read pid == 0 (treated as reclaimable), and win the reclaim CAS — tearing down a slot a live process just claimed and unlinking its in-use socket, so the endpoint silently vanishes from discovery.

Fix

• Claim the slot with an intermediate ENTRY_RESERVED sentinel that readers treat as empty.
• Write the payload under the seqlock, then publish the real state with a release store, so any reader observing the real state also observes the committed payload.
• Defense-in-depth: cleanup_stale() skips pid == 0 snapshots.

ENTRY_RESERVED lives only in the atomic state byte → no shm layout change, no REGISTRY_VERSION bump.

Files: registry.cpp, registry.hpp
Build: clean.

[benaliabderrahmane]

Reviewed the fix in depth -- it is a correct, surgical fix of a real cross-process P1 race, and I'm approving. I traced the memory ordering and it is sound:

• Writer does CAS state EMPTY -> RESERVED (acq_rel, registry.cpp:243-245), seq->odd, payload memcpy, seq->even (write_slot_payload), then state.store(type, release) at registry.cpp:250.
• A reader's acquire-load of state at registry.cpp:175 that observes the released type synchronizes-with the release store, so it also sees the committed payload and the even seq; a reader straddling publication gets s1 != s2 and retries.
• RESERVED is rejected at registry.cpp:175, so no half-built slot is ever snapshotted, and cleanup_stale (registry.cpp:380) only ever sees a real committed type as its CAS expected.
• ENTRY_RESERVED = 0xFF lives only in the atomic state byte (registry.hpp:35), so no shm-layout / REGISTRY_VERSION change is needed -- the static_asserts at registry.hpp:61-66 are unaffected. CI is green on jazzy/kilted/rolling.

Non-blocking follow-ups:

1. (CLAUDE.md sec.4) No reproducing/regression test ships with this fix. The commit touches only registry.cpp and registry.hpp -- 0 test files. A deterministic test looks feasible via the test-exposed registry_slots()/RegistryEntrySlot (cf. test/test_registry.cpp:163,177): write a non-empty state byte with seq left even and the payload zeroed, then assert snapshot_slot/registry_query returns nothing AND registry_cleanup_stale does not reclaim it (fails pre-fix, passes post-fix). The existing ReadersNeverSeeTornEntries (test_registry_concurrent.cpp:112-187) is probabilistic and never exercises cleanup_stale concurrently with adds, so it does not cover the reclaim-of-a-live-slot path.

2. (CLAUDE.md sec.3, orphan) registry.hpp:81 -- the reader-protocol comment step 2 still says if (st == EMPTY) skip, but snapshot_slot now also rejects ENTRY_RESERVED (registry.cpp:175). The writer steps were renumbered and the RESERVED enum was documented; the reader step was missed. One-line fix.

3. (nit, optional) The pid==0 guard in registry_cleanup_stale (registry.cpp:363-367) is redundant given the RESERVED fix and that all five production callers set pid=getpid(); it changes no behavior today (pid_is_alive_or_unreachable already rejects pid<=0). It's disclosed in the body as defense-in-depth and guards a catastrophic unlink, so I'm fine keeping it -- just noting the redundancy is intentional.

None of these block. Approving.