fix(v0.7-fire-tracker-4-final): widen env-scope keyword catch + anti-hallucination clause -- close all 7 LLM DM paths

Follow-up to v0.7-fire-tracker-4-revised. 7-path verify identified 4
paths failing for 2 root causes: (A) _ENV_KEYWORDS_TO_SUBTYPE missing
"traffic" + drop-audit phrases so build_traffic_detail and
build_drop_audit never ran; (B) LLM hallucinated specific numbers when
env blocks were empty (fabricated 144 earthquakes from a blank
quake_events table). This commit widens the keyword catch (phrase-match
for multi-word triggers to avoid false positives) and adds a positive-
framed anti-hallucination instruction to the env-context system prompt.
Re-verified all 7 paths against real Gemini in the prod container;
verdicts in v0.7-firetracker-phase4.md.

Class A -- routing miss fix:
- _ENV_KEYWORDS_TO_SUBTYPE gains "traffic"/"commute"/"highway"
  mapped to the existing "traffic" subtype.
- New _ENV_PHRASES_TO_SUBTYPE dict for multi-word triggers, matched
  as whole-phrase substrings (NOT single-word membership). Drop-audit
  phrases: "why didn't"/"why didnt"/"why am i not"/"why am i missing"
  /"what was filtered"/"drop audit"/"filtered out" all map to a new
  "drop_audit" subtype. Phrase-match keeps "why" alone from
  false-positing every "why is X" question.
- _detect_env_subtype now checks phrases first, then falls back to
  single-word tokenized match.

Class B -- positive-framed anti-hallucination clause:
- New module-level ENV_GROUNDING_CLAUSE constant. Appended to the
  system prompt whenever env scope is detected (after env_block +
  drop_block injection).
- Per Matt's mitigation guidance: positive ("answer from the blocks")
  not negative ("do not hallucinate"). Wording:
    "ENVIRONMENTAL CONTEXT GROUNDING:
     Answer only from the environmental context blocks above. If a
     block is empty or missing for an adapter the user asked about
     (e.g. no NWS alerts in the block), say something like 'No active
     <category> right now' -- never invent specific numbers, place
     names, or counts. If you do not have a relevant block for the
     question, say so briefly."

7-path verification, post-fix (real Gemini, prod container):

| # | query                                           | method               | verdict |
|---|-------------------------------------------------|----------------------|---------|
| 1 | "are there any fires near me?"                  | build_fires_detail   | PASS    |
| 2 | "any weather alerts?"                           | build_alerts_detail  | PASS    |
| 3 | "any earthquakes nearby?"                       | build_quakes_detail  | PASS    |
| 4 | "how's traffic on I-84?"                        | build_traffic_detail | PASS    |
| 5 | "what's the snake river level?"                 | build_gauges_detail  | PASS    |
| 6 | "what are the band conditions?"                 | build_swpc_detail    | PASS    |
| 7 | "why didn't I hear about anything today?"       | build_drop_audit     | PASS    |

Hallucination evidence (pre vs post on the quakes path):
  pre-fix: "There have been 144 earthquakes of magnitude 1.5 or
            greater in the past 24 hours worldwide. Some of the most
            recent earthquakes reported include: A magnitude 2.1
            earthquake in Pahala, Hawaii..."  (fabricated)
  post-fix: "I haven't observed any information about earthquakes
            in the mesh data."  (grounded)

Routing-miss evidence (pre vs post on the traffic path):
  pre-fix:  _detect_env_subtype("how's traffic on I-84?") -> None
            -> env scope NOT triggered, build_traffic_detail never
               called, LLM fabricated I-84 conditions in OR
  post-fix: _detect_env_subtype("how's traffic on I-84?") -> "traffic"
            -> env scope triggers, build_traffic_detail returns
               185 chars of real Ada-county incident data, LLM
               grounds on it: "I haven't observed any active
               traffic incidents on I-84 within the last two hours.
               The current active incidents are on North 9th Street
               / South 9th Street and SH-21, both in Ada."

Tests:
- 7 phase4 tests pass (no new tests needed; verification is the LLM
  DM path itself).
- Full suite: 56 passed in 3.80s across phase1+phase2+phase3+phase4
  +or-arch+include-roundtrip.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

meshai/router.py

@@ -91,18 +91,49 @@ _ENV_KEYWORDS_TO_SUBTYPE: dict[str, str] = {
     "road": "traffic", "roads": "traffic", "jam": "traffic",
     "crash": "traffic", "closure": "traffic", "511": "traffic",
     "incident": "traffic", "incidents": "traffic",
+    # v0.7-fire-4-final: "traffic"/"commute"/"highway" added so a
+    # query literally mentioning "traffic" hits the traffic subtype.
+    "traffic": "traffic", "commute": "traffic", "highway": "traffic",
     # generic
     "storm": "alerts", "weather": "alerts",
 }
 
 
-def _detect_env_subtype(message_lower: str) -> Optional[str]:
-    """Return the env subtype matched by the first env keyword in the message.
+# v0.7-fire-4-final: multi-word phrase triggers. Matched as whole-
+# phrase substrings of the message (NOT single-word membership) so
+# they carry multi-word semantics without a single word like "why"
+# or "filtered" firing false-positives in unrelated queries.
+# Drop-audit phrases unlock the env scope path so
+# env_reporter.build_drop_audit lands in the system prompt.
+_ENV_PHRASES_TO_SUBTYPE: dict[str, str] = {
+    "why didn't":       "drop_audit",
+    "why didnt":        "drop_audit",
+    "why am i not":     "drop_audit",
+    "why am i missing": "drop_audit",
+    "what was filtered":"drop_audit",
+    "drop audit":       "drop_audit",
+    "filtered out":     "drop_audit",
+}
 
-    `None` when no env keyword matches. Uses set intersection on tokenized
-    words so partial-word collisions (e.g. "firearm" / "fire") don\'t fire."""
+
+def _detect_env_subtype(message_lower: str) -> Optional[str]:
+    """Return the env subtype matched by the first env keyword/phrase
+    in the message. `None` when no env keyword matches.
+
+    v0.7-fire-4-final: phrase map is checked FIRST so multi-word
+    triggers (e.g. "why didn't I hear ...") work without their
+    constituent single words (e.g. "why" alone) firing false
+    positives. Single-word map then uses set-intersection on
+    tokenized words so partial-word collisions ("firearm" / "fire")
+    don't fire.
+    """
     if not message_lower:
         return None
+    # 1) Phrase substring match (multi-word semantics).
+    for phrase, subtype in _ENV_PHRASES_TO_SUBTYPE.items():
+        if phrase in message_lower:
+            return subtype
+    # 2) Single-word tokenized match.
     words = set(re.findall(r"\b\w+\b", message_lower))
     for kw, subtype in _ENV_KEYWORDS_TO_SUBTYPE.items():
         if kw in words:
@@ -777,6 +808,12 @@ class MessageRouter:
                 drop_block = env_reporter.build_drop_audit(hours=1)
                 if drop_block:
                     system_prompt += "\n\n" + drop_block
+                # v0.7-fire-4-final: positive-framed grounding clause.
+                # Closes Class B hallucination (LLM inventing counts
+                # / place names when an env block is empty -- e.g.
+                # "144 earthquakes worldwide" against an empty
+                # quake_events 24h window).
+                system_prompt += "\n\n" + ENV_GROUNDING_CLAUSE
             except Exception:
                 logger.exception("env_reporter injection failed")
 
@@ -949,3 +986,19 @@ class MessageRouter:
             connector=self.connector,
             history=self.history,
         )
+
+
+# v0.7-fire-4-final: positive-framed grounding clause appended to
+# the system prompt whenever env scope is detected. Frames the
+# constraint as "answer from the blocks" rather than "do not
+# hallucinate" (Matt's mitigation guidance) so the LLM doesn't
+# default to a blanket apology disclaimer every other message.
+ENV_GROUNDING_CLAUSE = (
+    "ENVIRONMENTAL CONTEXT GROUNDING:\n"
+    "Answer only from the environmental context blocks above. If a "
+    "block is empty or missing for an adapter the user asked about "
+    "(e.g. no NWS alerts in the block), say something like \"No "
+    "active <category> right now\" -- never invent specific numbers, "
+    "place names, or counts. If you do not have a relevant block for "
+    "the question, say so briefly."
+)