Correction: reject lost numbers and growth; place verifier anchors from the talkgroup name (server-26#162, server-26#160)

transcript_correction now also discards a correction that drops or changes a
bare number the recogniser heard (the live "7" -> "ShotSpotter" case) or
grows the transcript past max(4 words, 25%). place_verifier.verify() derives
an anchor from the talkgroup's own name plus the receiving node's state when
no area is configured; no state means no anchor, and a configured place
without an anchor keeps the load-bearing skip.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

drb-c2-core/app/internal/transcript_correction.py

@@ -116,11 +116,37 @@ def _dedupe(items: list[str]) -> list[str]:
 # mishearing this pass exists to fix (server-26#162).
 _CODE_TOKEN_RE = re.compile(r"\b\d{1,3}(?:-\d{1,3})+\b")
 
+# Bare numbers (the rest of server-26#162): the same call had a bare "7"
+# turned into "ShotSpotter". A unit number, a post, a house number or a count
+# the recogniser DID hear must survive, in order. A number may still be
+# added, because a sound-alike can legitimately be one ("back to Rose" ->
+# "back to Route 9"). "4th" is not bare (it is a word).
+_BARE_NUMBER_RE = re.compile(r"(?<![\w-])\d+(?![\w-])")
+
+# "NEVER add information" is the other prompt rule nothing checked. A
+# correction substitutes sound-alikes, so it may shift the word count a little
+# ("Holmes Street and 4th and Rowe" -> "Home Street and Forest Ave"); growing
+# by more than this is the model writing, not correcting.
+_MAX_ADDED_WORDS = 4
+_MAX_GROWTH_RATIO = 0.25
+
 
 def _code_tokens(text: str) -> list[str]:
     return _CODE_TOKEN_RE.findall(text or "")
 
 
+def _lost_numbers(original: str, corrected: str) -> bool:
+    """True unless every bare number of `original` is still in `corrected`, in order."""
+    remaining = iter(_BARE_NUMBER_RE.findall(corrected or ""))
+    return not all(any(n == m for m in remaining) for n in _BARE_NUMBER_RE.findall(original or ""))
+
+
+def _grew(original: str, corrected: str) -> bool:
+    n_orig = len((original or "").split())
+    n_new = len((corrected or "").split())
+    return n_new - n_orig > max(_MAX_ADDED_WORDS, int(n_orig * _MAX_GROWTH_RATIO))
+
+
 def _talkgroup_entry(system_doc: dict, talkgroup_id: Optional[int]) -> dict:
     """The config.talkgroups[] entry for this talkgroup, or {}."""
     if talkgroup_id is None:
@@ -241,6 +267,7 @@ async def correct(
     system_id: Optional[str] = None,
     talkgroup_id: Optional[int] = None,
     talkgroup_name: Optional[str] = None,
+    node_id: Optional[str] = None,
 ) -> tuple[Optional[str], Optional[list[dict]], bool]:
     """
     Second-opinion pass over a transcript.
@@ -330,6 +357,8 @@ async def correct(
                 context.get("tg_area"),
                 system_id=system_id,
                 talkgroup_id=talkgroup_id,
+                talkgroup_name=talkgroup_name,
+                node_id=node_id,
             )
         except Exception as e:
             logger.warning(f"Place verification failed for call {call_id}: {e}")
@@ -354,6 +383,18 @@ async def correct(
             f"discarding the joined correction"
         )
         corrected = None
+    if corrected is not None and _lost_numbers(text, corrected):
+        logger.warning(
+            f"Transcript correction for call {call_id} dropped or changed a number — "
+            f"discarding the joined correction"
+        )
+        corrected = None
+    if corrected is not None and _grew(text, corrected):
+        logger.warning(
+            f"Transcript correction for call {call_id} grew the transcript "
+            f"({len(text.split())} -> {len(corrected.split())} words) — discarding the joined correction"
+        )
+        corrected = None
     if corrected_segments is not None:
         for seg, orig in zip(corrected_segments, segments or []):
             if _code_tokens(seg["text"]) != _code_tokens(orig.get("text", "")):
@@ -363,6 +404,13 @@ async def correct(
                 )
                 corrected_segments = None
                 break
+            if _lost_numbers(orig.get("text", ""), seg["text"]) or _grew(orig.get("text", ""), seg["text"]):
+                logger.warning(
+                    f"Transcript correction for call {call_id} grew a segment or dropped "
+                    f"a number in one — discarding segment corrections"
+                )
+                corrected_segments = None
+                break
 
     if corrected or corrected_segments or not_speech:
         changed = raw.get("changed") or []