Match segment window by its own action phase

For segmented beats, the repair stage now searches for the source action
window using the segment's own description first; the full beat context is
used only as a fallback or when it scores noticeably higher. The trailer-
offset shift is applied only when the beat context is actually chosen.

Also harden vision-call retries to catch read-side network errors
(TimeoutError, socket.timeout, ConnectionError, OSError) and wrap the
filter/repair loop so a transient vision failure preserves the previously
cached match instead of dropping it.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/llm/vision_cache.py

@@ -13,6 +13,7 @@ import base64
 import json
 import logging
 import re
+import socket
 import time
 import urllib.error
 import urllib.request
@@ -193,6 +194,19 @@ def _call_vision_model(label: str, image_urls: list[str], cfg: AppConfig) -> str
                 len(delays_s),
             )
             time.sleep(delay_s)
+        except (TimeoutError, socket.timeout, ConnectionError, OSError) as exc:
+            if attempt >= len(delays_s):
+                raise RuntimeError(f"Vision request failed for {url}: {exc}") from exc
+            delay_s = delays_s[attempt]
+            logger.warning(
+                "Vision request network error for %s (%s); waiting %.0fs before retry %d/%d.",
+                label,
+                exc,
+                delay_s,
+                attempt + 1,
+                len(delays_s),
+            )
+            time.sleep(delay_s)
 
     raise RuntimeError(f"Vision request failed unexpectedly for {url}")