fix: error state, leg grouping, prompt→generating transition

- log_parser: detect model load failure (child + router patterns) → state=error;
  switch to generating at progress=1.00 and at prompt eval summary line
- frontend: add state-error (llama flashes red, CSS variable override);
  group each leg+hoof in <g> so walk animation moves hooves with legs

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

log_parser.py

@@ -83,6 +83,8 @@ _PROMPT_SUMMARY_RE = re.compile(
 )
 _RELEASE_RE = re.compile(r'slot\s+release:.*stop processing')
 _UNLOAD_RE = re.compile(r'unload_lru:|unload: stopping model instance')
+_LOAD_ERROR_RE = re.compile(r'exiting due to model loading error')
+_EXIT_ERROR_RE = re.compile(r'operator\(\):.*exited with status 1')
 
 
 # ── State machine ─────────────────────────────────────────────────────────────
@@ -121,6 +123,8 @@ class LogParser:
             return self._set(state="waiting")
         if _UNLOAD_RE.search(content):
             return self._set(state="unloading")
+        if _EXIT_ERROR_RE.search(content):
+            return self._set(state="error")
         return None
 
     # ── Child-process lines ──────────────────────────────────────────────────
@@ -130,6 +134,8 @@ class LogParser:
 
         if m := _CMD_STATE_RE.match(content):
             return self._handle_cmd_state(m.group(1))
+        if _LOAD_ERROR_RE.search(content):
+            return self._set(state="error")
         if _WARMUP_RE.search(content):
             return self._set(state="warming_up")
         if _IDLE_RE.search(content):
@@ -137,14 +143,14 @@ class LogParser:
         if _LAUNCH_RE.search(content):
             return self._set(state="processing_prompt")
         if m := _PROMPT_TIMING_RE.search(content):
+            progress = float(m.group(2))
+            done = progress >= 1.0
             metrics = Metrics(
                 prompt_speed=float(m.group(3)),
                 prompt_tokens=int(m.group(1)),
-                prompt_progress=float(m.group(2)),
-                gen_speed=self._state.metrics.gen_speed,
-                n_decoded=self._state.metrics.n_decoded,
+                prompt_progress=None if done else progress,
             )
-            return self._set(state="processing_prompt", metrics=metrics)
+            return self._set(state="generating" if done else "processing_prompt", metrics=metrics)
         if m := _GEN_TIMING_RE.search(content):
             metrics = Metrics(
                 prompt_speed=self._state.metrics.prompt_speed,
@@ -154,14 +160,18 @@ class LogParser:
             )
             return self._set(state="generating", metrics=metrics)
         if m := _PROMPT_SUMMARY_RE.search(content):
-            # End-of-slot summary: update prompt speed without changing state.
-            # For short prompts this fires after generating starts (no live prompt line).
             metrics = Metrics(
                 prompt_speed=float(m.group(2)),
                 prompt_tokens=int(m.group(1)),
                 gen_speed=self._state.metrics.gen_speed,
                 n_decoded=self._state.metrics.n_decoded,
             )
+            if self._state.state == "processing_prompt":
+                # Summary fires right after prompt eval finishes, before first token.
+                # Advance to generating now rather than waiting for the first tg line.
+                # For short prompts the summary arrives after generating has already
+                # started, so the guard prevents stepping back.
+                return self._set(state="generating", metrics=metrics)
             return self._set(metrics=metrics)
         if _RELEASE_RE.search(content):
             return self._set(