fix(doctor): name the processes holding a retired WAL generation instead of a green state.db

After the deleted-WAL guard fires, every SessionDB open raises
DeletedWalGenerationError, yet `hermes doctor` printed "✓ state.db exists"
and "0 process(es) holding the DB open" with rc 0 (#110054). It only counted
holders of the CURRENT state.db inode; the processes wedging the store are
the ones holding the RETIRED -wal/-shm inodes, and nothing in-product named
them — while the chat explainer sends users to exactly this doctor.

Ask iter_deleted_sqlite_sidecar_holders (the helper the guard itself uses)
first. When it names holders: warn with their PIDs, record a stop-them issue
line pinned to the active profile, and return before the health/stats/WAL
probes — a read-only connect on the new generation is another opener, and
--fix must not checkpoint under them. No kill path, no new module.

hermes_cli/doctor_state.py

@@ -344,11 +344,32 @@ def _state_db_wal(f: Finding, should_fix: bool, state_db_path: Path) -> None:
             check_info(f"WAL file is {size // (1024*1024)} MB (normal for active sessions)")
 
 
+def _retired_wal_holders(f: Finding, state_db_path: Path, _DHH: str) -> bool:
+    """Name the processes holding a retired -wal/-shm generation (#110054). Every SessionDB open is
+    refused while they live, and the current inode has no holders, so the plain holder count says
+    "0 holding the DB open" beside a green state.db line — the opposite of the truth."""
+    from hermes_constants import profile_cli_selector
+    from hermes_state_dbfile import iter_deleted_sqlite_sidecar_holders
+    pids = list(dict.fromkeys(pid for pid, _ in iter_deleted_sqlite_sidecar_holders(state_db_path)))
+    if not pids:
+        return False
+    rendered = ", ".join(f"pid {pid}" for pid in pids)
+    check_warn(f"{_DHH}/state.db: {len(pids)} process(es) still hold a retired WAL generation ({rendered})",
+               "(every new session refuses to open until they exit; health/stats probes skipped)")
+    f.issues.append(f"state.db retired WAL generation held by {rendered} — stop them: "
+                    f"'hermes {profile_cli_selector()}gateway stop', quit the Desktop app / dashboard, "
+                    "then rerun 'hermes doctor'")
+    return True
+
+
 @doctor_check()
 def _check_state_db(should_fix: bool, f: Finding) -> None:
     """state.db session count, FTS write health, schema repair, stats snapshot, WAL size."""
     from hermes_cli.doctor import HERMES_HOME, _DHH
     state_db_path = HERMES_HOME / "state.db"
+    # A read-only connect on the new generation is itself another opener, so nothing below may run.
+    if _retired_wal_holders(f, state_db_path, _DHH):
+        return
     if state_db_path.exists():
         _state_db_health(f, should_fix, state_db_path, _DHH)
         _state_db_stats(f.issues, state_db_path)

tests/hermes_cli/test_doctor_wal_holder_guard.py

@@ -49,3 +49,29 @@ def test_wal_checkpoint_skipped_while_live_writer_holds_db(tmp_path):
 
     assert finding.fixed == 0
     assert any("gateway" in issue for issue in finding.issues)
+
+
+def test_doctor_names_retired_wal_holders_instead_of_healthy_state_db(tmp_path, monkeypatch, capsys):
+    """After the deleted-WAL guard fires (#110054), doctor must name the PIDs holding the retired
+    generation, must not print a healthy state.db line, and must not open the store itself (the
+    health probe is another opener) nor checkpoint under --fix."""
+    import hermes_cli.doctor as doctor
+    import hermes_cli.doctor_state as doctor_state
+    import hermes_state_dbfile
+
+    db = tmp_path / "state.db"
+    db.write_bytes(b"")
+    monkeypatch.setattr(doctor, "HERMES_HOME", tmp_path)
+    monkeypatch.setattr(hermes_state_dbfile, "iter_deleted_sqlite_sidecar_holders",
+                        lambda path: [(4242, f"{path}-wal"), (4242, f"{path}-shm")])
+    probed = []
+    monkeypatch.setattr(doctor_state, "_state_db_health", lambda *a, **k: probed.append(a))
+    monkeypatch.setattr(doctor_state, "_state_db_wal", lambda *a, **k: probed.append(a))
+
+    finding = doctor_state._check_state_db(True)
+    out = capsys.readouterr().out
+
+    assert "4242" in out and "retired WAL" in out
+    assert "✓" not in out
+    assert probed == [] and finding.fixed == 0
+    assert any("4242" in issue and "gateway stop" in issue for issue in finding.issues)