Branch and bound

.gitignore

@@ -1,6 +1,11 @@
 __pycache__/
 .DS_Store
 **/.DS_Store
+.env
+*.zip
+__MACOSX/
+**/__MACOSX/
+*.egg-info/
 .venv/
 run_exp.sh
 eval_exp.sh
@@ -16,3 +21,8 @@ uv.lock
 .venv
 agent_env/runs
 agent_env/config.toml
+rank_cache/
+.claude/
+LISTEN_v1/
+LISTEN_v2/
+LISTEN_v2 2/

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "listen"]
+	path = listen
+	url = git@github.com:AdamJovine/LISTEN-active.git
+	branch = china

BNB_EXPERIMENT_LOG.md

@@ -0,0 +1,151 @@
+# BnB Experiment Log — Session 2026-06-17
+
+## Objective
+
+Maximize pass rate on 95 originally-failing queries while keeping 39-passing sample at 100%.
+All runs used `--rank-cache rank_cache` (no LLM calls, pure cache replay).
+
+---
+
+## Run Index
+
+### Failing-set runs (95 queries originally failing at baseline)
+
+| Commit | Run dir | Pass | Total | Rate | Notes |
+|--------|---------|------|-------|------|-------|
+| `6dfbe0a` | `results/failures_at_6dfbe0a/run_20260617_192902_bnb_groq` | 70 | 86 | 81% | Baseline (only 86 ran; 9 had errors/timeouts) |
+| `04e2793` | `results/failures_at_04e2793/run_20260617_211323_bnb_groq` | 77 | 94 | 82% | After unuse_restaurant fix |
+| `094f087` | `results/failures_at_094f087/run_20260617_213504_bnb_groq` | 80 | 95 | 84% | After poi_ordering fix |
+
+### Passing-sample runs (39 queries that pass at baseline — regression guard)
+
+| Commit | Run dir | Pass | Total | Rate |
+|--------|---------|------|-------|------|
+| `6dfbe0a` | `results/passing_sample_at_6dfbe0a/run_20260617_193914_bnb_groq` | 39 | 39 | 100% |
+| `04e2793` | `results/passing_sample_assembler_fix/run_20260617_205749_bnb_groq` | 39 | 39 | 100% |
+| `094f087` | `results/passing_sample_poiorder_fix/run_20260617_213300_bnb_groq` | 39 | 39 | 100% |
+
+### Full-1000 runs
+
+| Run dir | Pass | Total | Rate | Notes |
+|---------|------|-------|------|-------|
+| `results/FULL1000_BNB_MAX/run_20260616_184445_bnb_groq` | 905 | 1000 | 90.5% | Full dataset before this session's fixes |
+| `results/run_20260617_180524_bnb_groq` | 132 | 152 | 86.8% | Partial run during session |
+
+---
+
+## Changes and Impact
+
+### `04e2793` — assembler: unuse_restaurant + opentime check
+
+**+1% on failing set** (70→77/94, 81%→82%)
+
+**Root cause:** `next_restaurant()` marks a restaurant as "used" before `_meal_activity()` checks
+if it can actually fit. If a restaurant (e.g. Chef's Wife, opens 11:00) was selected for a
+lunch slot where arrival was after 14:00 (the lunch window close), it was permanently consumed
+and unavailable for dinner.
+
+**Fix 1:** Skip restaurants where `opentime >= meal_window_close` in `next_restaurant()`.
+Added `_MEAL_WIN_CLOSE = {"breakfast": "09:00", "lunch": "14:00", "dinner": "20:00"}` dict
+and pre-check in all three loop paths (priority, pool order, fallback).
+
+**Fix 2:** Added `unuse_restaurant()` closure after `next_restaurant()` that returns a
+restaurant to the pool (by object identity) when `_meal_activity()` or `_can_afford()` rejects
+it. Added `else: unuse_restaurant(rest)` at all 6 meal-slot call sites in assembler.py.
+
+**File:** `chinatravel_tpc/assembler.py`
+
+---
+
+### `094f087` — poi_ordering: ordered pin list + attr_pins re-sort
+
+**+2% on failing set** (77→80/95, 82%→84%)
+
+**Root cause 1:** `_parse_pinned_names()` used Python `set`, losing insertion order. For queries
+requiring "Manchester United Dream Theater then Houhai", the set could return Houhai first,
+prepending it before Manchester United in `attr_pins` and causing the poi_ordering constraint
+to fail.
+
+**Fix 1:** Changed `_parse_pinned_names()` to use ordered `list` (preserving insertion order
+from nl2sl literals). `pins.setdefault(cat, [])` with dedup check.
+
+**Fix 2:** Added explicit poi_ordering-aware `attr_pins` re-sort in `bnb_agent.py`. Parses
+`idx_activity` patterns from `hard_logic_py` snippets to extract (before_name, after_name)
+pairs, then swaps `attr_pins` entries that are in the wrong order.
+
+**Files:** `chinatravel_tpc/agent.py`, `chinatravel_tpc/bnb_agent.py`
+
+---
+
+### `ab8f858` — agent: normalize interior whitespace in pin names
+
+**Impact not yet measured** (committed 2026-06-17, run pending)
+
+**Root cause:** Some DB hotel names have double spaces (e.g., "Atour X Hotel Shanghai  Hongqiao
+Airport Konggang Road") but nl2sl predicts single space. `_normalize_pin_name` only handled
+parenthesis spacing.
+
+**Fix:** Added `re.sub(r'\s+', ' ', name)` before the parenthesis normalization to collapse
+all interior whitespace.
+
+**Expected impact:** Resolves 1 of the 3 `required_hotel_name` failures (UID 20250322205724212077,
+Atour X Hotel case). The other two hotel cases appear to be a different bug where the pinned
+hotel is found by `_rows_for_pins` but never selected in the BnB skeleton phase.
+
+**File:** `chinatravel_tpc/agent.py`
+
+---
+
+## Remaining 15 Failures at `094f087`
+
+Run: `results/failures_at_094f087/run_20260617_213504_bnb_groq`
+
+| UID | Failure type | Notes |
+|-----|-------------|-------|
+| 20250321040138918100 | `other` | Unknown constraint |
+| 20250321114239878527 | `required_attraction_type` | |
+| 20250322130400129262 | `required_hotel_feature` | |
+| 20250322142408120417 | `OR_compound` | Hangzhou 4-day, sightseeing ≤200 OR single bed |
+| 20250322161842269069 | `required_cuisine_type` | Japanese cuisine at airport hotel area |
+| 20250322164349699070 | `required_restaurant` | |
+| 20250322165301153800 | `required_cuisine_type` | Japanese cuisine at airport hotel area |
+| 20250322171831366188 | `other` | |
+| 20250322192555845893 | `poi_timing` | Peppa Pig 12:30-14:00 window |
+| 20250322194001155137 | `required_hotel_name` | "Shanghai Lujiazui Babaiban Lan'ou Hotel" — pin found but skeleton ignores it |
+| 20250322205724212077 | `required_hotel_name` | "Atour X Hotel Shanghai Hongqiao" — double-space mismatch (fixed in ab8f858) |
+| 20250323010327713880 | `poi_timing` | Bistro Sola 17:00-18:00 window |
+| 20250323031255302334 | `total_budget`, `poi_timing` | Blue Airflow Skydiving 14:20-15:50 window |
+| 20250324082922869744 | `required_attraction` | |
+| 20250324223916776179 | `required_hotel_name` | Third hotel pin case |
+
+### Open investigations
+
+**Hotel pin not selected in skeleton (2 cases: 20250322194001155137, 20250324223916776179):**
+- `_rows_for_pins(hotel_df_full, ...)` finds the hotel (verified)
+- `_prepend_pins(hotel_pins, hotel_winners)` puts it first in hotel_winners
+- But skeleton output shows only other hotels (Waiting Hotel, Jinglai Hotel)
+- Likely cause: `_find_feasible_skeletons` evaluates hotel against `hard_logic` and the
+  hotel fails some constraint (proximity? budget?), or `hotel_winners` is being re-sorted
+  after `_prepend_pins` runs
+- Next step: add debug print inside `_find_feasible_skeletons` to see why pinned hotel is
+  rejected in the skeleton phase
+
+**poi_timing (3 cases):**
+- Assembler visits attraction outside its open window
+- Need to check if `min_visit_duration` is consuming too much time before the timed attraction
+
+**required_cuisine_type (2 cases):**
+- Japanese restaurants near airport hotel areas not being selected
+- May be a rank cache issue (ranked low) or proximity filter excluding them
+
+---
+
+## What Remains To Do
+
+1. **Run failing set at `ab8f858`** to confirm double-space fix resolves UID 20250322205724212077
+2. **Investigate hotel pin skeleton bypass** — why do pinned hotels get prepended to
+   `hotel_winners` but never appear in feasible skeletons (UIDs 20250322194001155137,
+   20250324223916776179)
+3. **Investigate poi_timing failures** (3 cases) — assembler timing logic
+4. **Investigate required_cuisine_type failures** (2 cases) — Japanese cuisine availability
+5. **Run full 1000 at latest commit** after fixes accumulate to get updated full-dataset score