Call-1 reasoning budget 1024 + no-think A/B flag; narration name lockdown

Live trace findings drove these:
- Call-1 reasoning ran past the 512-token budget without emitting JSON
(empty -> fallback). Bump think budget to 1024; no-think stays 256.
- Add DISCOVERROUTE_VIBE_THINKING A/B flag (config.VIBE_THINKING, default
on) and record the chosen mode on every vibe_extraction trace row so
think vs no-think can be compared head-to-head.
- Narration kept name-dropping real off-route landmarks (Pont des Arts,
Mona Lisa, Sorbonne, Sacre-Coeur) -> gate rejected -> template. Tighten
the prompt to a hard allowlist (stops + endpoints + gazetteer only;
scene-set without proper names) and fix the literal-"H3" artifact.
- Scale narration token budget with route length (600..880) so long
routes stop truncating mid-name.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/discoverroute/config.py

@@ -149,6 +149,16 @@ EMBED_QUERY_INSTRUCTION = "Represent this sentence for searching relevant passag
 # LlamaForCausalLM architecture — no custom kernels.
 LLM_MODEL = "openbmb/MiniCPM5-1B"
 
+# A/B toggle for Call 1 (vibe→weights): run the model's REASONING pass
+# (enable_thinking, MiniCPM5-1B is hybrid-reasoning) or the fast no-think path.
+# Thinking may score a fuzzy vibe better but needs a far bigger token budget and
+# more of the ZeroGPU slice. Flip the Space variable DISCOVERROUTE_VIBE_THINKING
+# (1=think [default], 0=no-think); the chosen mode is recorded on every trace row
+# so the two can be compared head-to-head. (Either way, degenerate/empty output
+# falls through to the bge-small embed tier — the route is never left tasteless.)
+VIBE_THINKING = os.environ.get(
+    "DISCOVERROUTE_VIBE_THINKING", "1").lower() in ("1", "true", "on")
+
 # --- Trace logging (Open Trace) ----------------------------------------------
 # Every inference call logs a row locally to logs/traces.jsonl; when a write
 # token is present, rows are ALSO pushed (async, non-blocking) to TRACE_REPO.

src/discoverroute/interpret/llm_vibe.py

@@ -121,19 +121,22 @@ def extract(vibe: str) -> dict | None:
         {"role": "system", "content": SYSTEM_PROMPT},
         {"role": "user", "content": f"Vibe: {vibe}\nReturn the JSON weights now."},
     ]
+    from discoverroute import config
+    think = config.VIBE_THINKING
+    meta_in = {"vibe": vibe, "thinking": think}  # mode recorded for the A/B
     t0 = time.time()
     try:
         from discoverroute.narrate.llm import run_inference
-        # Reasoning pass: MiniCPM5-1B is hybrid-reasoning, and scoring a fuzzy vibe
-        # across 14 types is exactly the kind of short deliberation a 1B does better
-        # with than off-the-cuff. enable_thinking=True lets it reason, then emit the
-        # JSON; run_inference strips the <think> block and returns only the JSON. The
-        # budget covers the reasoning + the ~120-token object (truncated reasoning →
-        # empty answer → clean fallback).
-        raw_text = run_inference(messages, max_new_tokens=512, enable_thinking=True)
+        # A/B (config.VIBE_THINKING): with thinking ON, MiniCPM5-1B reasons before
+        # emitting JSON — but the <think> block + the ~120-token object need real
+        # room, so give it 1024 (512 ran past the budget and returned empty in the
+        # first live test). No-think is lean and fast. run_inference strips the
+        # <think> block; a truncated/unclosed reasoning → empty answer → fallback.
+        budget = 1024 if think else 256
+        raw_text = run_inference(messages, max_new_tokens=budget, enable_thinking=think)
     except Exception as exc:  # noqa: BLE001 - never break interpretation
         latency = int((time.time() - t0) * 1000)
-        trace.log_trace("vibe_extraction", {"vibe": vibe},
+        trace.log_trace("vibe_extraction", meta_in,
                         {"error": f"{type(exc).__name__}: {exc}"},
                         latency, used_fallback=True)
         return None
@@ -143,13 +146,13 @@ def extract(vibe: str) -> dict | None:
     # Reject unparseable OR degenerate (all-zero / all-equal) output: both leave the
     # router with no taste signal, so fall through to the embedding tier instead.
     if parsed is None or _is_degenerate(parsed):
-        trace.log_trace("vibe_extraction", {"vibe": vibe},
+        trace.log_trace("vibe_extraction", meta_in,
                         {"raw": raw_text, "degenerate": parsed is not None},
                         latency, used_fallback=True)
         return None
 
     affinity = mapping.brief_scores_to_affinity(parsed)
     budget_mult = max(0.5, min(2.0, parsed["detour_budget_multiplier"]))
-    trace.log_trace("vibe_extraction", {"vibe": vibe}, parsed,
+    trace.log_trace("vibe_extraction", meta_in, parsed,
                     latency, used_fallback=False)
     return {"affinity": affinity, "budget_multiplier": budget_mult, "raw": parsed}

src/discoverroute/narrate/narrate.py

@@ -254,24 +254,27 @@ def _llm_narration(plain, discovery, pois, vibe, mode, start_label, end_label,
     system = (
         f"You are a {guide}local — a sharp, warm city guide who actually walks these "
         "streets. Write a vivid, first-person walk for this route.\n"
-        "FORMAT — this matters: the first line is EXACTLY ONE markdown H3 heading — a "
-        "short, evocative title (3–6 words) you invent for THIS specific walk, hinting "
-        "at its mood and place. After that, write 2–4 flowing paragraphs "
-        "of continuous prose. Do NOT number the stops. Do NOT give any stop its own "
-        "heading or bullet. If there are many stops, weave several into each paragraph "
-        "rather than a sentence each — keep it moving and finish before you run long. "
-        "Carry the reader start→finish with sensory detail and natural transitions "
-        "('a block on', 'just around the corner', 'as the street opens up'). Bold each "
-        "real stop's name the first time it appears, and reference the user's vibe in "
-        "your own words.\n"
+        "FORMAT — this matters: the FIRST line is a single short title (3–6 words) you "
+        "invent for THIS walk, written as a markdown H3 — it MUST begin with '### ' "
+        "(three hashes then a space). Never write the literal letters 'H3'. After the "
+        "title, write 2–4 flowing paragraphs of continuous prose. Do NOT number the "
+        "stops. Do NOT give any stop its own heading or bullet. If there are many "
+        "stops, weave several into each paragraph rather than a sentence each — keep it "
+        "moving and finish before you run long. Carry the reader start→finish with "
+        "sensory detail and natural transitions ('a block on', 'just around the "
+        "corner', 'as the street opens up'). Bold each real stop's name the first time "
+        "it appears, and reference the user's vibe in your own words.\n"
         "Describe each place in your OWN words — say what it is and why it's worth it. "
         "Never write bare type labels like 'park garden', 'place of worship' or "
-        "'monument historic'. Set the scene freely with the districts, river and "
-        "landmarks under 'You may reference', and mention the time of day or the light.\n"
-        "ONE hard rule: do not invent a *named venue to visit* — every place you name "
-        "as a stop must be one of the 'Ordered stops' (spelled close to the list) or "
-        "the start/destination. You don't need a name in every sentence; describe "
-        "freely, just never fabricate a café/shop/museum name that isn't on the list."
+        "'monument historic'. Mention the time of day or the light.\n"
+        "ONE HARD RULE — NAMES: the ONLY proper names you may write are the 'Ordered "
+        "stops', the start/destination, and the items under 'You may reference'. Do NOT "
+        "name ANY other place — no landmark, museum, monument, bridge, street, square, "
+        "university, church or artwork, even a real famous one (for example do not name "
+        "the Mona Lisa, Sacré-Cœur, Pont des Arts, the Sorbonne, the Louvre) unless it "
+        "is in those lists. To set the scene, describe WITHOUT proper names "
+        "('a stone bridge', 'a grand old church'). A single name outside the lists "
+        "makes the entire itinerary get thrown away."
     )
     user = (
         f"Vibe: {vibe or 'open to anything'}\n"
@@ -285,10 +288,13 @@ def _llm_narration(plain, discovery, pois, vibe, mode, start_label, end_label,
     )
     messages = [{"role": "system", "content": system},
                 {"role": "user", "content": user}]
-    # 600 tokens lets flowing prose cover ~10-12 stops without cutting off mid-walk
-    # ("### 9: End your trip" in the traces); a 1B model on A10G still finishes inside
-    # the 45s ZeroGPU slice (see llm.GPU_DURATION_S).
-    text = run_inference(messages, max_new_tokens=600)
+    # Scale the budget with route length so long walks finish instead of truncating
+    # mid-name (a live trace cut "Sacré-C[œur]" off at 600 → the gate then rejected
+    # the fragment). Short routes stay lean; long ones get up to 880, still inside the
+    # 45s ZeroGPU slice (see llm.GPU_DURATION_S). No-think path (fast, direct prose).
+    n = len(pois)
+    max_tok = 600 if n <= 6 else min(880, 600 + 40 * (n - 6))
+    text = run_inference(messages, max_new_tokens=max_tok)
     # Title guard: a 1B model sometimes parrots a stock example title or omits the
     # heading. Ensure the walk opens with a sensible H3 — swap in our own computed
     # route-title if the model echoed a known example or wrote no heading at all.