Round-4 fixes: drop overclaiming match-tag, dedup stops, hidden excludes famous

Adversarial round 4 (NOT-CONVERGED) — root-caused to 3 issues:
1. The per-stop "a match for your vibe" tag overclaimed: any route that
backfills a lower-ranked category (bakery on wine, church on jazz, statue
on street-food/photography) falsely labelled those stops a match. Removed
the tag entirely — each stop's reason text + the interpretation panel carry
the honest framing; the model can't guarantee per-stop relevance.
2. No within-route dedup → the same place ("UP /SIDE / DOWN / TOWN") appeared
twice in one route. Dedup the shortlist by osm_id AND name.
3. "hidden gems" still surfaced Notre Dame — the attraction-zeroing missed it
(it's viewpoint). Discovery cues now drop famous, well-documented POIs by
tag-richness (confidence >= FAMOUS_CONFIDENCE 0.85), category-independent.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/discoverroute/config.py

@@ -111,6 +111,10 @@ TOP_AFFINITY_CATEGORIES = 6
 # architecture" 0.51, nonsense ~0.49). We still route, but the narration says so
 # honestly instead of claiming "a match for your vibe".
 WEAK_MATCH_SIMILARITY = 0.55
+# For "hidden gems"-style vibes, exclude well-documented (famous) POIs: a place
+# this richly tagged isn't off the beaten path. Confidence is the tag-richness
+# proxy; Notre Dame etc. sit at ~1.0. (Only applied when a discovery cue fires.)
+FAMOUS_CONFIDENCE = 0.85
 # Below this cosine-similarity span across categories, a vibe is treated as
 # off-domain/neutral rather than amplified into false preferences. Measured
 # (bge-small, 16-vibe battery): gibberish "asdfqwer" spans 0.081; the LOWEST

src/discoverroute/interpret/vibe.py

@@ -48,6 +48,7 @@ class Interpretation:
     confidence: float = 1.0        # best raw cosine to a category gloss
     weak: bool = False             # True => out-of-vocabulary / weak match
     adventurousness: float = config.DEFAULT_ADVENTUROUSNESS  # possibly cue-boosted
+    exclude_famous: bool = False   # discovery cue => drop well-documented sights
 
 
 def _contains(text: str, cues) -> bool:
@@ -63,7 +64,8 @@ def interpret(vibe: str, adventurousness: float = config.DEFAULT_ADVENTUROUSNESS
     # Discovery intent: drop "famous attraction" (opposite of "hidden") and lift
     # adventurousness so the route favours under-documented places. Copy first —
     # resolve_affinity is cached and returns a shared dict.
-    if _contains(text, _HIDDEN_CUES):
+    hidden = _contains(text, _HIDDEN_CUES)
+    if hidden:
         affinity = {**affinity, "attraction": 0.0}
         adventurousness = max(adventurousness, 0.8)
     weights = Weights(category_affinity=affinity, w_category=1.0)
@@ -95,7 +97,7 @@ def interpret(vibe: str, adventurousness: float = config.DEFAULT_ADVENTUROUSNESS
     explanation = _explain(vibe, top, affinity, posture, budget_hint, weak)
     return Interpretation(affinity, weights, posture, budget_hint, explanation, top,
                           confidence=confidence, weak=weak,
-                          adventurousness=adventurousness)
+                          adventurousness=adventurousness, exclude_famous=hidden)
 
 
 def _explain(vibe, top, affinity, posture, budget_hint, weak=False) -> str:

src/discoverroute/narrate/narrate.py

@@ -79,13 +79,12 @@ def template_narration(plain, discovery, pois, vibe, mode, start_label="",
         f"**{n} {place_word}** between {start_label or 'the start'} and "
         f"{end_label or 'the destination'}:\n"
     )
-    # The categories the vibe leans on most — used to tie a stop back to the vibe.
-    # A weak vibe has no real matches, so don't tag stops "a match for your vibe".
-    top_cats: set[str] = set()
-    aff = getattr(weights, "category_affinity", None)
-    if v and aff and not weak:
-        top_cats = set(sorted(aff, key=aff.get, reverse=True)[:3])
-
+    # NOTE: we deliberately do NOT tag stops "a match for your vibe". Adversarial
+    # review showed that claim overreaches whenever a route backfills with a
+    # lower-ranked category (a bakery on a wine vibe, a church on a jazz vibe) —
+    # the affinity model can't guarantee per-stop relevance, so the blanket claim
+    # reads as a lie. Each stop's own reason text conveys its appeal honestly; the
+    # interpretation panel already shows how the vibe was read.
     lines = [lead]
     prev_cat = None
     for i, p in enumerate(pois, 1):
@@ -96,7 +95,7 @@ def template_narration(plain, discovery, pois, vibe, mode, start_label="",
             reason = _REASON.get(p.category, "a spot worth a look")
         prev_cat = p.category
         verb = _verb(posture.get(p.category, "pass"))
-        tie = " — a match for your vibe" if (v and p.category in top_cats) else ""
+        tie = ""
         badge = _hours_badge(p, posture.get(p.category, "pass"))
         lines.append(f"{i}. **{label}** — {verb.lower()} for {reason}{tie}.{badge}")
     lines.append(

src/discoverroute/pipeline.py

@@ -115,6 +115,7 @@ def _plan_route_impl(
     from discoverroute.data import taxonomy
     interp_md = ""
     weak_match = False
+    exclude_famous = False  # discovery-cue vibes drop well-documented famous sights
     top_requested = None  # the #1 category the vibe asked for (for sparse feedback)
     posture = {c: taxonomy.posture(c) for c in taxonomy.CATEGORIES}
     has_vibe = bool((vibe or "").strip())
@@ -131,6 +132,7 @@ def _plan_route_impl(
             interp_md = interp.explanation
             weak_match = interp.weak
             adventurousness = interp.adventurousness  # may be cue-boosted (e.g. "hidden gems")
+            exclude_famous = interp.exclude_famous
             if not weak_match and interp.top_categories:
                 top_requested = interp.top_categories[0]
             # Use the interpreter's OWN affinity (it carries discovery-cue
@@ -168,7 +170,7 @@ def _plan_route_impl(
     try:
         shortlist, matrix, time_fn = _prepare_discovery(
             graph, start, end, plain, mode, budget, weights, adventurousness,
-            posture=posture)
+            posture=posture, exclude_famous=exclude_famous)
         for _ in range(max(1, n_alternatives)):
             if shortlist is None:
                 break
@@ -227,7 +229,7 @@ def _plan_route_impl(
 
 
 def _prepare_discovery(graph, start, end, plain, mode, budget, weights, adventurousness,
-                       posture=None):
+                       posture=None, exclude_famous=False):
     """Corridor → score → shortlist → real travel matrix. Done ONCE per request.
 
     The expensive step is the matrix (cutoff-bounded multi-source Dijkstra), so we
@@ -238,13 +240,34 @@ def _prepare_discovery(graph, start, end, plain, mode, budget, weights, adventur
     candidates = poimod.corridor_pois(plain.coords, budget)
     if not candidates:
         return None, None, None
+    # Discovery vibes ("hidden gems"): drop famous, well-documented sights so the
+    # route stays genuinely off the beaten path (Notre Dame etc. enter via several
+    # categories, so filter by tag-richness, not category).
+    if exclude_famous:
+        kept = [p for p in candidates if p.confidence < config.FAMOUS_CONFIDENCE]
+        if kept:
+            candidates = kept
     scoring.score_pois(candidates, weights, adventurousness)
     # Open-now awareness: demote places that are closed right now (heavily for
     # stop-at categories, mildly for pass-by; unknown hours left untouched).
     from discoverroute.routing import hours
     hours.apply_open_now(candidates, posture)
-    shortlist = sorted((p for p in candidates if p.score > 0),
-                       key=lambda p: p.score, reverse=True)[: config.SOLVER_CANDIDATES]
+    ranked = sorted((p for p in candidates if p.score > 0),
+                    key=lambda p: p.score, reverse=True)
+    # Dedup within a route: the same OSM place can appear as multiple rows
+    # (multipolygon centroids) or two distinct ids can share a name — either way
+    # a route must never tell you to visit the same spot twice. Keep first (best).
+    seen_id, seen_name, shortlist = set(), set(), []
+    for p in ranked:
+        nkey = (p.name or "").strip().lower()
+        if p.osm_id in seen_id or (nkey and nkey in seen_name):
+            continue
+        seen_id.add(p.osm_id)
+        if nkey:
+            seen_name.add(nkey)
+        shortlist.append(p)
+        if len(shortlist) >= config.SOLVER_CANDIDATES:
+            break
     if not shortlist:
         return None, None, None