fix: make sim_warp_factor real-hardware-safe and sud logs forecast-reproducible

sim_warp_factor was scaling DT_TASK unconditionally, including against real
hardware, and was leaking into recorded sample timestamps via wall-clock
time - both violate its intended meaning (a reciprocal scale on task wait
time only). Clamp it to 1.0 whenever Heater.type isn't "sim", and track
elapsed time via tick-counted accumulators (Sud.elapsed for SudLogTask, a
matching one for ServerLogTask) instead of time.monotonic().

Sud/server logs now also carry HeaterPowers, the effective SimWarpFactor,
the raw Doc, a correct Name, and ForecastAnchors (one entry per real
_reanchor_forecast() trigger) - enough for utils/analyze_log.py to rebuild
a SudForecastEstimator and replay the exact same splice-per-transition
forecast correction offline that a live client sees, replacing the old
(always-dead) forecast_*.json sidecar lookup.

Two bugs found and fixed along the way:
- ServerLogTask's periodic wall-clock checkpoint had no way to know a
  SudLogTask run had just ended, so it would immediately re-write the
  just-completed log with PlantParams/ForecastAnchors cleared back to
  empty. Gated the checkpoint on _should_sample().
- _reanchor_forecast() truncated the old forecast against real elapsed
  time, which can end up numerically less than the old forecast's own
  (very wrong) speculative timestamps once a WAIT_USER confirmation runs
  long - leaving a "ghost" segment instead of a clean cut. Truncate
  against forecast_step_starts[index] instead, which lives in the same
  coordinate space as the forecast being cut.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JFDeoTKBDkXRhPfnyDEoBt

tasks/sud.py

@@ -113,6 +113,7 @@ class SudTask(ATask):
 		self._on_end = None
 		self._on_start = None
 		self._on_plant_params = None
+		self._on_reanchor = None
 		msg_handler.set_recv_handler(self.recv)
 
 	def set_on_plant_params(self, callback):
@@ -120,6 +121,16 @@ class SudTask(ATask):
 		so external observers (e.g. SudLogTask) can record each change."""
 		self._on_plant_params = callback
 
+	def set_on_reanchor(self, callback):
+		"""Register a callback invoked with (elapsed, index, theta_ist)
+		every time _reanchor_forecast() fires (see on_step_changed()) -
+		lets an external observer (SudLogTask) record the exact anchor
+		points a live client's forecast got corrected at, so an offline
+		re-simulation (utils/analyze_log.py) can reproduce the same
+		splice-per-transition forecast instead of just the single,
+		uncorrected cold-start one."""
+		self._on_reanchor = callback
+
 	def apply_plant_params(self, grain_mass, water_mass):
 		"""Keeps the real plant's and the controller's internal model's
 		plant params in sync with this Sud's own doc - L/Td come straight
@@ -215,6 +226,12 @@ class SudTask(ATask):
 			# (a malt fill-in's actual cooldown, a longer/shorter ramp than
 			# modeled, ...) at the next opportunity, not just at the next
 			# user confirmation.
+			if self._on_reanchor:
+				# Read synchronously, right here - the same values
+				# _reanchor_forecast() itself will read moments later, off
+				# the same unyielded call stack, before anything else can
+				# mutate them.
+				self._on_reanchor(self.sud.elapsed, self.sud.index, self.tc.get_theta_ist())
 			asyncio.create_task(self._reanchor_forecast())
 
 		asyncio.create_task(self.send({'Step': {
@@ -373,16 +390,30 @@ class SudTask(ATask):
 		self._forecast_generation += 1
 		generation = self._forecast_generation
 		real_elapsed = self.sud.elapsed
+		schedule = self.sud.schedule
+		index = self.sud.index
 		# Drop the now-stale tail (everything beyond right now) - it's
-		# about to be replaced by a freshly anchored simulation.
-		cut = bisect.bisect_right(self.forecast_t, real_elapsed)
+		# about to be replaced by a freshly anchored simulation. Cut at
+		# forecast_step_starts[index], the old forecast's own (possibly
+		# very wrong) belief of where step `index` begins - not at
+		# real_elapsed directly: a step whose real timing blew way past
+		# what its zero-delay guess assumed (a long WAIT_USER confirm,
+		# above all - see SudForecastEstimator.estimate()'s docstring)
+		# leaves the old forecast's *entire* speculative remainder sitting
+		# at timestamps still numerically less than real_elapsed, so
+		# bisecting against real_elapsed itself would find nothing to
+		# discard and just tack the fresh, correct simulation on after it
+		# - a doubled-back, self-overlapping curve instead of a clean cut.
+		# forecast_step_starts[index] doesn't have this problem: it lives
+		# in the same (speculative) coordinate space as forecast_t itself,
+		# so the cut lands in the right place regardless of how far real
+		# and speculated time have diverged by now.
+		cut = bisect.bisect_right(self.forecast_t, self.forecast_step_starts.get(index, real_elapsed))
 		forecast_t = self.forecast_t[:cut]
 		forecast_theta = self.forecast_theta[:cut]
 		# Steps already passed (< index) have their real, now-immutable
 		# start time; anything from index on is about to be resimulated
 		# fresh below and must not keep a stale prediction around.
-		schedule = self.sud.schedule
-		index = self.sud.index
 		forecast_step_starts = {i: tt for i, tt in self.forecast_step_starts.items() if i < index}
 
 		if not (0 <= index < len(schedule)):