diff --git a/NOTEPAD.md b/NOTEPAD.md
index a2737f8..8bb9732 100644
--- a/NOTEPAD.md
+++ b/NOTEPAD.md
@@ -12,6 +12,29 @@ Deferring because out of scope is fine, but don't mention deferring than.
 
 ### To be sorted
 
+- BUG/FLOW: **a cancelled request can leave `FlowController._lock` held forever.**
+  Seeding nineteen flows over a client that timed out mid-request left the next
+  `POST /flows/{name}/start` waiting on the lock indefinitely — ten minutes, until
+  the container was restarted. A py-spy dump showed *no* thread in the reload path,
+  so the coroutine that holds it is suspended at an `await` inside `reload()`, most
+  likely in `_teardown()` awaiting a supervised task's cancellation. Everything else
+  kept working — health, reads, MQTT — so the engine looked fine and only anything
+  needing a rebuild hung. Two things worth doing: release the lock on cancellation
+  (`asyncio.timeout` around the teardown, or a `finally` that cannot be skipped), and
+  fail a `start` that waits more than a few seconds for the lock rather than hanging.
+- BUG/INFRA: **475 zombie `git` processes** in the API container after a seeding
+  session. `FlowStore._git` uses `subprocess.run`, which reaps its own child — these
+  are the `git gc --auto` daemons `git commit` spawns, reparented to PID 1 when their
+  parent exits. PID 1 is the FastAPI process, which never reaps orphans. Harmless at
+  484 processes against a 34719 limit, but it grows with every save, and a container
+  that commits per keystroke will get there. Fix: `git -c gc.auto=0 commit`, since a
+  store that never packs is a separate (and real) concern — 1898 commits had left
+  5863 loose objects and no packs.
+- PERF/API: seeding nineteen flows takes two reloads each — one to publish, one to
+  stop — so a full rebuild of every flow in the installation runs about forty times
+  for one seed. It is the slowest thing about standing an installation up, and a
+  `PUT` that could say "published, stopped" in one call would halve it.
+
 - FEAT/SEC: `locked` is a read-only surface, not a permission — the server accepts a publish
   from a panel whose dashboard says locked. Making it real means carrying the flag into
   `_panel_may`.