docs: correct os.replace claim and complete the hand-rolled site list

Phase 2 review findings: (1) _commit_staged_replacements' docstring cited
os.replace while the code uses os.rename — the atomicity claim holds (same-
filesystem rename is atomic on POSIX and NTFS) but named the wrong function.
(2) venv_bin_dir's remaining hand-rolled site list missed agent/lsp/servers.py:270.

hermes_cli/update_cmd.py

@@ -665,9 +665,10 @@ def _commit_staged_replacements(staged) -> None:
     This covers plain files as well as directories: the repo root holds 20
     first-party modules (``run_agent.py``, ``cli.py``, ``hermes_constants.py``
     …), so a files-only failure reproduces exactly the bug class we are
-    closing. ``os.replace`` is atomic on POSIX and maps to
-    ``MoveFileEx(REPLACE_EXISTING)`` on Windows, so a file swap can never
-    leave a half-written module the way ``copy2`` onto a live path can.
+    closing. Every swap is an ``os.rename`` onto a path that was just moved
+    aside — a same-filesystem rename is atomic on POSIX and NTFS alike, so a
+    file swap can never leave a half-written module the way ``copy2`` onto a
+    live path can.
 
     Splitting stage-all-then-swap-all shrinks the failure window from "the
     duration of a full tree copy" to "the duration of N renames", and makes

hermes_constants.py

@@ -1261,8 +1261,8 @@ def venv_bin_dir(venv_dir, *, windows: bool | None = None) -> Path:
     each new call site had to re-derive it, and #76091 shipped an eighth copy
     because the correct behaviour lived 2400 lines away in another function.
     A few sites outside ``hermes_cli`` (``tools/code_execution_tool.py``,
-    ``agent/lsp/install.py``) still hand-roll it — convert them as they are
-    touched.
+    ``agent/lsp/install.py``, ``agent/lsp/servers.py``) still hand-roll it —
+    convert them as they are touched.
 
     *windows* lets a caller pass its own platform verdict. Several callers
     resolve this through predicates the test-suite patches to exercise