diff --git a/components/TODO.md b/components/TODO.md
new file mode 100644
index 0000000..a0d9f53
--- /dev/null
+++ b/components/TODO.md
@@ -0,0 +1,58 @@
+# components/sud.py step lifecycle spec
+
+Confirmed spec for a `Sud` schedule step's internal phase order
+(`components/sud.py`), from a design discussion clarifying how `ramp`/
+`hold`/user-confirm interact.
+
+## Phase order
+
+`ramp` -> `hold` -> user-confirm. Each step independently opts into
+`ramp` and/or `hold` by including that key (`_build_step()`,
+`sud.py:35-45`); the other phase is skipped entirely if its key is
+absent. The `'ramp'` key's presence is the *only* trigger for ramping
+(`_advance()`, `sud.py:192-210`) - there is no diff-based decision
+(`T_actual` vs. `T_target`) anywhere that decides whether to ramp, and
+no internal setpoint-trajectory limiter in the temperature controller
+(`set_theta_soll()` is an instant, unconditional assignment,
+`components/pid/temp_controller_base.py:81-82`). The `ramp` key both
+gates ramping *and* carries its settings (`rate`); it does not merely
+override settings on top of an always-on ramp.
+
+## Duration `0`
+
+Supported for hold: `hold_remaining = duration * 60.0` can be `0`;
+`tick()` (`sud.py:186-190`) finishes the step on the next tick once
+`hold_remaining <= 0`.
+
+## End-of-step branching
+
+`_finish_step()` (`sud.py:212-216`): no `user_wait_for_continue` ->
+advances immediately to the next step; `user_wait_for_continue: true`
+-> enters `WAIT_USER`, resumed only by `confirm()`.
+
+`WAIT_USER` is not "hold, prolonged": `HOLDING` and `WAIT_USER` are
+distinct `SudState` values. The hold countdown runs to completion
+*before* `WAIT_USER` begins - `tick()` only decrements
+`hold_remaining` while `state == HOLDING`, so the timer is frozen
+(not extended) during the wait. It delays the next step, but it's a
+separate third phase, not a longer hold.
+
+The gating field is `user_wait_for_continue`, not "presence of a user
+message": `user_message` (display text) and `user_wait_for_continue`
+(the actual block) are independent fields - a step can show a message
+without blocking, or block silently.
+
+## Open items
+
+- [ ] **Pure-hold steps never (re-)set the controller's target
+  temperature.** For a step with only `hold` (no `ramp`),
+  `tasks/sud.py`'s `on_step_changed` never calls `tc.set_theta_soll()`
+  - the guard at line 73 requires `ramp is not None`. The
+  controller's target temperature for a pure-hold step is whatever
+  was left from the previous step's ramp; there's no explicit
+  fallback to the hold step's own `temperature` field. `README.md`
+  (Mash schedules section) documents this as designed behavior
+  ("`SudTask` drives ... `theta_soll`/`heatrate_soll` from ramp
+  steps ... counts down hold steps' duration"), so it's intentional,
+  not a bug - but it means a schedule can't express a temperature
+  jump straight into a hold without a `ramp` block.