Refresh narrative docs

docs/DATAANALYSIS.md

@@ -0,0 +1,344 @@
+# DATAANALYSIS — what the curated data actually says
+
+This document is the analytical companion to the curated parquet bundle
+produced by `geography-vn curate`. It walks through the 14-figure pack
+under [`docs/figures/analysis/`](docs/figures/analysis/) (rendered by
+`scripts/analyze.py` and embedded inline in
+[`DATAANALYSIS.ipynb`](DATAANALYSIS.ipynb)) and tells the story of
+**Vietnam after Resolution 202/2025/QH15** through the data alone.
+
+> **Source:** every number below comes from
+> `data/sapnhap-bando-vn/extracted/extracted.parquet` (6,712 rows = 34
+> provinces + 3,321 communes + 3,357 commune people's-committee
+> headquarters). The parquet was produced by the five-stage NeMo Curator
+> pipeline described in [`DATAPROCESSING.md`](DATAPROCESSING.md).
+
+## TL;DR
+
+| Stat                                  | Value             |
+| ------------------------------------- | ----------------- |
+| First-level units (post-merger)       | **34**            |
+| Second-level units (post-merger)      | **3,321**         |
+| People's-committee headquarters       | **3,357**         |
+| Total Vietnam population (2024)       | **113,571,926**   |
+| Total Vietnam land area (km²)         | **331,325.62**    |
+| Largest province by population        | TPHCM — **14,002,598**   |
+| Largest province by area              | Lâm Đồng — **24,233 km²** |
+| Densest province                      | Hà Nội — **2,621 / km²**  |
+| Most-merged commune                   | Phường Văn Miếu - Quốc Tử Giám — **16 predecessor wards** |
+| Distinct authorising decrees          | **37**            |
+| Provinces that kept their old borders | **11** out of 34  |
+
+## 1 — Inventory ([fig 01](docs/figures/analysis/01_admin_kind_donut.png))
+
+The atlas publishes three entity kinds, in roughly equal sub-counts:
+
+| Kind        | Count   | What it is                                              |
+| ----------- | ------- | ------------------------------------------------------- |
+| `province`   | 34      | First-level admin units (28 provinces + 6 cities)        |
+| `commune`    | 3,321   | Second-level units (phường / xã / đặc khu after the merger) |
+| `committee`  | 3,357   | Commune people's-committee headquarters (point markers) |
+
+The committee count slightly exceeds the commune count (3,357 > 3,321)
+because some communes have multiple registered committee buildings (the
+seat plus a satellite office) and a handful of pre-merger committees
+remain marked on the map even after their parent commune dissolved.
+
+## 2 — Macro-region balance ([fig 02](docs/figures/analysis/02_macro_region_breakdown.png))
+
+The 34 surviving provinces redistribute across the six GSO macro-regions
+unevenly. The north (where the 2025 merger was the most aggressive) keeps
+**10 + 5 = 15 provinces** under two macro-regions; the centre keeps
+**8 + 3 = 11**; the south keeps **3 + 5 = 8**. Including the
+people's-committee tier (which inherits its parent province's region):
+
+| Macro-region (EN)                          | Provinces | Communes | Committees | Total |
+| ------------------------------------------ | --------: | -------: | ---------: | ----: |
+| Northern Midlands and Mountain Areas       | **10**    |   841    |    842     | 1,693 |
+| Red River Delta                            | **5**     |   527    |    502     | 1,034 |
+| North Central and Central Coastal Areas    | **8**     |   738    |    737     | 1,483 |
+| Central Highlands                          | **3**     |   361    |    380     |   744 |
+| Southeast                                  | **3**     |   359    |    358     |   720 |
+| Mekong River Delta                         | **5**     |   495    |    537     | 1,037 |
+| (unknown — disputed-zone duplicates)       |    0      |     0    |      1     |     1 |
+
+Note that the **Northern Midlands** retains the highest *count* of
+first-level units (10) despite being the region the merger consolidated
+hardest — pre-merger it had 14 provinces. The **Southeast** ends up the
+smallest by province count (3) because TPHCM swallowed both Bình Dương
+and Bà Rịa-Vũng Tàu into a single mega-city.
+
+## 3 — Province population ([fig 03](docs/figures/analysis/03_province_population.png))
+
+The top-five most populous provinces, after the merger:
+
+| Rank | Province                | Population   | Area (km²) | Density (/km²) | Predecessors |
+| ---- | ----------------------- | -----------: | ---------: | -------------: | -----------: |
+| 1    | Thành Phố Hồ Chí Minh    | 14,002,598   |  6,772.59  | 2,067.5        | 3            |
+| 2    | Thủ Đô Hà Nội           |  8,807,523   |  3,359.84  | 2,621.4        | 1            |
+| 3    | Tỉnh An Giang            |  4,952,238   |  9,888.91  |   500.8        | 2            |
+| 4    | Thành Phố Hải Phòng      |  4,664,124   |  3,194.72  | 1,460.0        | 2            |
+| 5    | Thành phố Đồng Nai      |  4,491,408   | 12,737.18  |   352.6        | 4            |
+
+TPHCM's merger absorbed **Bình Dương** (industrial belt) and **Bà Rịa
+- Vũng Tàu** (the southern coast & Côn Đảo) into one super-city, doubling
+its population from ~7 M to **14 M** — now larger than New York's metro.
+
+Đồng Nai is the only province in the country that absorbed **four**
+predecessor units (the original Đồng Nai + Bình Phước + parts of two
+neighbours), making it the largest-area populated unit in the south.
+
+## 4 — Province area ([fig 04](docs/figures/analysis/04_province_area.png))
+
+The top-five largest provinces by km², all in the central highlands or
+upper-central coast:
+
+| Rank | Province          | Area (km²)  | Population | Density (/km²) | Predecessors |
+| ---- | ----------------- | ----------: | ---------: | -------------: | -----------: |
+| 1    | Tỉnh Lâm Đồng    | 24,233.07   |  3,872,999 | 159.8           | 3            |
+| 2    | Tỉnh Gia Lai      | 21,576.53   |  3,583,693 | 166.1           | 2            |
+| 3    | Tỉnh Đắk Lắk      | 18,096.40   |  3,346,853 | 184.9           | 2            |
+| 4    | Tỉnh Nghệ An      | 16,486.50   |  3,831,694 | 232.4           | 1            |
+| 5    | Tỉnh Quảng Ngãi   | 14,832.55   |  2,161,755 | 145.7           | 2            |
+
+Lâm Đồng is the merger's largest single creation: the old Lâm Đồng +
+**Bình Thuận** (south-central coast) + **Đắk Nông** (Central Highlands)
+fused into a 24,233 km² province that now stretches from the central
+plateau to the South China Sea coast.
+
+## 5 — Province density ([fig 05](docs/figures/analysis/05_province_density.png))
+
+| Rank | Province                | Density (/km²) | Population | Area (km²) |
+| ---- | ----------------------- | -------------: | ---------: | ---------: |
+| 1    | Thủ Đô Hà Nội           | 2,621.4        |  8,807,523 |  3,359.84  |
+| 2    | Thành Phố Hồ Chí Minh    | 2,067.5        | 14,002,598 |  6,772.59  |
+| 3    | Thành Phố Hải Phòng      | 1,460.0        |  4,664,124 |  3,194.72  |
+| 4    | Tỉnh Hưng Yên            | 1,418.8        |  3,567,943 |  2,514.81  |
+| 5    | Tỉnh Ninh Bình           | 1,119.1        |  4,412,264 |  3,942.62  |
+
+Hà Nội remains the densest province even though it kept its pre-merger
+borders unchanged (its `n_predecessors == 1`). Hưng Yên, **post-merger**
+absorbing Thái Bình, jumps into the top 4 — the merged Red River Delta
+provinces concentrate density there.
+
+## 6 — Communes per province ([fig 06](docs/figures/analysis/06_communes_per_province.png))
+
+The lower-tier consolidation is more uneven:
+
+* **Min**: 38 communes (Tỉnh Lai Châu — sparsely populated mountains)
+* **Max**: 168 communes (Thành phố Hồ Chí Minh — mega-city)
+* **Median**: 99 communes per province
+* **Mean**: 97.7 communes per province
+
+Bottom five (sparsest):
+
+| Province           | Communes |
+| ------------------ | -------: |
+| Tỉnh Lai Châu      |    38    |
+| Thành phố Huế      |    40    |
+| Tỉnh Điện Biên     |    45    |
+| Tỉnh Quảng Ninh    |    54    |
+| Tỉnh Cao Bằng      |    56    |
+
+Top five (densest):
+
+| Province                   | Communes |
+| -------------------------- | -------: |
+| Thành phố Hồ Chí Minh      |   168    |
+| Tỉnh Thanh Hóa             |   166    |
+| Tỉnh Phú Thọ               |   148    |
+| Tỉnh Gia Lai               |   135    |
+| Tỉnh Nghệ An               |   130    |
+
+## 7 — Province merger fanout ([fig 07](docs/figures/analysis/07_merger_fanout_provinces.png))
+
+How many predecessor provinces fed each surviving first-level unit:
+
+| n_predecessors | Provinces | Names                                                                                                    |
+| -------------- | --------: | -------------------------------------------------------------------------------------------------------- |
+| **1**           | **11**    | Hà Nội, Huế, Cao Bằng, Điện Biên, Hà Tĩnh, Lai Châu, Lạng Sơn, Nghệ An, Quảng Ninh, Sơn La, Thanh Hóa     |
+| **2**           | **16**    | Đà Nẵng, Hải Phòng, An Giang, Bắc Ninh, Cà Mau, Đắk Lắk, Đồng Tháp, Gia Lai, Hưng Yên, Khánh Hòa, … |
+| **3**           | **6**     | Cần Thơ, Hồ Chí Minh, Lâm Đồng, Ninh Bình, Phú Thọ, Vĩnh Long                                          |
+| **4**           | **1**     | Đồng Nai                                                                                                  |
+
+Eleven provinces kept their old borders — they were already large enough
+or geographically isolated enough that the merger left them alone. The
+modal merger absorbed exactly one neighbour. Đồng Nai is the singular
+four-way merger.
+
+## 8 — Commune merger fanout ([fig 08](docs/figures/analysis/08_merger_fanout_communes.png))
+
+The lower tier was consolidated much more aggressively. Of the 3,321
+surviving communes:
+
+| n_predecessors | Communes | %     |
+| -------------- | -------: | ----: |
+|  **1**          |    139   |  4.2% |
+|  **2**          |    758   | 22.8% |
+|  **3**          |  1,498   | 45.1% |
+|  **4**          |    559   | 16.8% |
+|  **5**          |    179   |  5.4% |
+|  **6**          |     78   |  2.3% |
+|  **7+**         |    110   |  3.4% |
+| **max**         |     16   |       |
+
+The modal commune absorbed **3 predecessor wards/xã**. The mean fanout
+is 3.4 (compared to ~1.6 for the first-level tier).
+
+The **most-merged commune is Phường Văn Miếu - Quốc Tử Giám** in Hà
+Nội, which fused **16 predecessor wards** into a single 4-character
+neighbourhood. Hà Nội dominates the high-fanout extreme — 8 of the top
+15 communes by fanout are in Hà Nội, reflecting the post-merger desire
+to flatten the dense old ward grid into larger, more administratively
+manageable units. The other notable mega-mergers are coastal special
+zones: **Đặc khu Cát Hải** (Hải Phòng, 12 preds), **Đặc khu Vân Đồn**
+(Quảng Ninh, 12 preds).
+
+## 9 — Commune size distribution ([fig 09](docs/figures/analysis/09_commune_size_distribution.png))
+
+A log-log scatter of commune area vs population reveals three regimes:
+
+* **Special administrative zones (Đặc khu)** — sit in the bottom-left:
+  tiny populations on tiny islands. Hoàng Sa (the Paracels — 0
+  registered population, 350 km²), Trường Sa (Spratly — 153 people,
+  496 km²), Cồn Cỏ (139 people, 2.3 km²), Bạch Long Vĩ (686 people,
+  3.07 km²). All have ambiguous-to-disputed sovereignty status and
+  serve mainly as markers of national territory.
+* **Highland xã** — the upper-left band: tens of thousands of
+  inhabitants spread across 200-1100 km² of mountain terrain. Buôn Đôn
+  (Đắk Lắk, 1,114 km², 6.6K people) is the largest by area in the entire
+  set. The Trường Sơn cordillera communes in Quảng Trị (Thượng Trạch,
+  Trường Sơn, Kim Ngân) all exceed 800 km² each.
+* **Urban phường** — the lower-right cluster: 50-130K people on a few
+  km² of city. Phường Hải Châu (Đà Nẵng, 7.58 km², 131K people) is the
+  densest by far.
+
+## 10 — Decree corpus ([fig 10](docs/figures/analysis/10_decree_map.png))
+
+The merger is authorised by **37 distinct decrees**: 1 from the National
+Assembly itself (`Nghị quyết số 202/2025/QH15`) and 34 follow-up
+Standing Committee resolutions (`NQ-UBTVQH15`) of 16 June 2025 that flesh
+out the lower-tier merger commune-by-commune. The top decrees by number
+of units cited:
+
+| Decree                            | Units cited |
+| --------------------------------- | ----------: |
+| `Nghị quyết số 1685/NQ-UBTVQH15`   | 168          |
+| `Nghị quyết số 1686/NQ-UBTVQH15`   | 166          |
+| `Nghị quyết số 1676/NQ-UBTVQH15`   | 148          |
+| `Nghị quyết số 1664/NQ-UBTVQH15`   | 135          |
+| `Nghị quyết số 1678/NQ-UBTVQH15`   | 130          |
+| `Nghị quyết số 1674/NQ-UBTVQH15`   | 129          |
+| `Nghị quyết số 1656/NQ-UBTVQH15`   | 126          |
+| … (30 more)                       | …            |
+
+Each `NQ-1656…1690` decree corresponds to one of the 34 surviving
+provinces. The top decree (NQ-1685) authorises all 168 wards of TPHCM;
+NQ-1686 covers the 166 communes of Thanh Hóa; etc. Every commune row in
+the parquet has a `decree_url` column pointing to the official text at
+`vanban.chinhphu.vn`.
+
+## 11 — Cartographic pack with dual-archipelago declaration
+
+Geographic visualisations live in their own folder under
+[`docs/figures/maps/`](docs/figures/maps/) — rendered standalone by
+`python -m scripts.render_maps` because kaleido + Mapbox-GL crashes
+inside a Jupyter kernel. Five figures total, each with the **actual
+scraped province polygons** (not bubbles) and the standard
+Vietnamese-atlas declaration of both offshore archipelagos:
+* **Quần đảo Hoàng Sa** (Paracel Islands) — dashed bounding outline
+  ~15.45°N–17.20°N × 110.85°E–113.10°E with markers for Phú Lâm, Tri Tôn,
+  Linh Côn, Quang Hòa. Administered by **Thành phố Đà Nẵng** under the
+  post-2025 geography (Đặc khu Hoàng Sa).
+* **Quần đảo Trường Sa** (Spratly Islands) — dashed bounding outline
+  ~7.40°N–11.85°N × 111.80°E–115.30°E with markers for Trường Sa Lớn,
+  Song Tử Tây, Sinh Tồn, Phan Vinh, An Bang, Nam Yết, Cô Lin, Sơn Ca.
+  Administered by **Tỉnh Khánh Hòa** (Đặc khu Trường Sa).
+
+Each map additionally carries: ★ Thủ đô Hà Nội (capital), 7 secondary
+cities (TP. Hồ Chí Minh / Đà Nẵng / Hải Phòng / Huế / Vinh / Cần Thơ /
+Nha Trang) with leader-line labels into open sea, and 5 island callouts
+(Đảo Phú Quốc / Cát Bà / Bạch Long Vĩ / Lý Sơn + Côn Đảo).
+
+| File                                            | Visualisation                                                |
+| ----------------------------------------------- | ------------------------------------------------------------ |
+| `maps/01_provinces_population.png`               | population choropleth (sequential NVIDIA-green)              |
+| `maps/02_provinces_density.png`                  | people / km²                                                 |
+| `maps/03_provinces_area.png`                     | land area (km²)                                              |
+| `maps/04_communes_scatter.png`                   | 3,321 commune centroids by macro-region                      |
+| `maps/05_committees_scatter.png`                 | 3,357 commune people's-committee headquarters by lat/lon     |
+
+The cartographic pack uses NVIDIA Sans typography on a square 1100 × 1100
+canvas; the analytical pack above uses the LaTeX-serif `nvidia_latex`
+template at 1200 × 900. Maps follow brand convention, charts follow
+academic-paper convention.
+
+## 12 — Curator UMAP ([fig 11](docs/figures/analysis/11_curator_umap_kind.png), [fig 12](docs/figures/analysis/12_curator_umap_region.png))
+
+A 2-D UMAP projection of the 6,712-point sentence-transformers
+embedding of every entity's `embed_text` descriptor. Three observations:
+
+1. **Kind separation is partial.** Provinces and committees occupy
+   compact, well-separated tail regions of the map (provinces
+   because their descriptors include rich predecessor-province prose
+   that other entities do not; committees because every committee
+   description follows the standard "Ủy ban nhân dân …" template).
+   Communes fill the middle and overlap with both extremes.
+2. **Macro-region structure emerges.** Even though the embed text
+   does not explicitly mention macro-region, the UMAP × macro-region
+   colouring shows clean lobes for the Mekong Delta, the Northern
+   Midlands, and the Central Highlands — the model picks up on
+   regional naming conventions (`Đặc khu Phú Quốc` clusters with the
+   southern coast, `Xã Mường Khương` with the Northwest, etc.).
+3. **HDBSCAN finds ~30-50 dense subclusters** within the broader
+   regional structure. Most are sub-province-level: a single
+   province's communes cluster together because their descriptors
+   share a common `parent_ten` tag, the same parent decree, and
+   similar predecessor-name patterns ("Phường … và Phường … sau khi
+   sắp xếp"). The `cluster` column in `reduced.parquet` carries the
+   integer label per row, with `-1` reserved for low-density noise
+   points.
+
+## Caveats & known limits
+
+* **`Đặc khu Hoàng Sa` has zero registered population.** The Paracel
+  Islands are administered by the People's Republic of China and have
+  not been physically Vietnamese-controlled since 1974. The Vietnamese
+  atlas marks them as a special administrative zone of Đà Nẵng with a
+  populations of zero — a sovereignty assertion, not a usable
+  demographic figure.
+* **Population numbers are 2024 mid-year estimates** as published with
+  the merger decrees (rounded thousands at the commune level). They
+  predate the merger by 6–12 months and may not yet reflect intra-merger
+  population migration.
+* **Two parent-province strings come back inconsistent:** "Thủ đô Hà Nội"
+  vs "Thủ Đô Hà Nội" (different `Đ`/`đ` capitalisation depending on the
+  endpoint). The `parent_ma` column normalises this — join on `parent_ma`
+  not on `parent_ten` for analysis.
+* **3 communes carry the name "Xã Hoàng Hoa Thám"** (each in a different
+  province) and 1 carries "Đặc khu Phú Quốc". Use `id` (== feature id)
+  not `ten` as the primary key.
+* **The committee macro-region** is derived from the GeoJSON
+  `properties.a04_tentinh` field, which can be missing for ~20 of the
+  3,357 committees (mostly defunct pre-merger seats). Those rows show
+  `macro_region == "unknown"` in the parquet.
+
+## How to reproduce
+
+```bash
+# 1) Download + curate (~35 min on first run, then cached)
+geography-vn curate
+
+# 2) Render the figure pack (~2 min)
+python -m scripts.analyze
+
+# 3) Re-execute the notebook (optional)
+jupyter nbconvert --to notebook --execute --inplace DATAANALYSIS.ipynb
+
+# 4) Stage and push the HF dataset
+python -m scripts.upload_to_hf --no-upload
+python -m scripts.upload_to_hf --repo <your-org>/sapnhap-bando-vn
+```
+
+See [`DATAPROCESSING.md`](DATAPROCESSING.md) for the curator-pipeline
+internals.

docs/DATAPROCESSING.md

@@ -0,0 +1,227 @@
+# DATAPROCESSING — five-stage curator pipeline
+
+This document walks through the five-stage NeMo-Curator-compatible pipeline
+that turns the raw web responses from <https://sapnhap.bando.com.vn/> into
+the typed, joinable, embedding-augmented parquet bundle that ships to
+HuggingFace and feeds [`DATAANALYSIS.md`](DATAANALYSIS.md).
+
+The five stages mirror the same layout used by ViLA's
+[`packages/datasites/anle/`](https://github.com/tmquan/ViLA/tree/main/packages/datasites/anle)
+and `personas-vn`'s
+[`packages/curator/`](https://github.com/tmquan/personas-vn/tree/main/packages/curator):
+
+```
+download → parse → extract → embed → reduce
+```
+
+Each stage reads from the previous stage's output directory and writes to its
+own. A failure inside one stage never destroys earlier work, and re-running a
+single stage in isolation (`--only embed`, `--skip download`, …) is the
+intended development loop.
+
+```
+data/sapnhap-bando-vn/
+├── raw/         ← stage 1
+│   ├── admin_units.json          (one POST: 3,355 listings)
+│   ├── committees.json           (one POST: 3,357 committee markers)
+│   ├── details/<malk>.json       (~3,355 POSTs to p.co_dvhc_id)
+│   ├── geom/<id>.geojson         (~6,712 POSTs to pread_json)
+│   └── _cache/                   (per-URL HTTP cache; re-runs are free)
+├── parsed/      ← stage 2: parsed.jsonl + parsed.parquet
+├── extracted/   ← stage 3: extracted.jsonl + extracted.parquet
+├── embedded/    ← stage 4: embedded.parquet (n × 384-d vectors)
+└── reduced/     ← stage 5: reduced.parquet (UMAP 2-D + cluster id)
+```
+
+## Source surface
+
+The site at `sapnhap.bando.com.vn/` is a thin PHP front-end
+(`D:\map34tinh\s.index.php`) over a QGIS Server WMS/WFS deployment. It
+exposes four POST endpoints we care about — every one returns JSON, even
+when the server lies in its `Content-Type` header:
+
+| Endpoint                | Form data                       | Returns                                              |
+| ----------------------- | ------------------------------- | ---------------------------------------------------- |
+| `POST /p.co_dvhc`        | `ma=0`                           | List of every admin unit (34 prov + 3,321 communes)  |
+| `POST /p.co_uyban`       | `ma=0`                           | List of 3,357 commune people's-committee headquarters |
+| `POST /p.co_dvhc_id`     | `malk=<feature_id>`              | Full attribute row (area, population, decree, …)     |
+| `POST /pread_json`       | `id=<feature_id>`                | GeoJSON FeatureCollection (Polygon / MultiPolygon / Point) |
+
+Feature-id conventions:
+
+* `diaphanhanhchinhcaptinh_sn.<n>`   — province polygons (only 34 of the 132
+  pre-merger ids survive).
+* `diaphanhanhchinhcapxa_2025.<n>`  — commune polygons; 3,321 alive, the
+  rest dissolved into neighbours.
+* `uybannhandancapxa_2025.<n>`     — point markers for every commune
+  people's committee (n = 1 … 3,357).
+
+The PHP front-end occasionally injects an HTML warning preamble before the
+JSON body when QGIS Server is mid-restart — `packages.common.http` peels
+that off transparently.
+
+## Stage 1 — download
+
+```python
+DownloadStage(config.download, raw_dir).run()
+```
+
+* **Two listing POSTs** capture the complete inventory in one shot each
+  (`/p.co_dvhc` and `/p.co_uyban`).
+* **Per-unit detail walk** (~3,355 POSTs to `/p.co_dvhc_id`) pulls the rich
+  attribute row for every admin unit: area in km², population, capital,
+  predecessors prose, decree of authority, link to the official decree at
+  `vanban.chinhphu.vn`.
+* **Per-feature geometry walk** (~6,712 POSTs to `/pread_json`) pulls the
+  polygon for every admin unit and the point marker for every committee.
+* Every URL is cached on disk under `raw/_cache/`; re-runs hit the cache and
+  finish in seconds.
+* `delay_between_requests_s: 0.10` keeps the crawl polite — ~10 req/s, no
+  hint of rate limiting from the server.
+* Wall time on first run: ~12 minutes for the listings + details, plus
+  ~22 minutes for the geometries; ~35 minutes end-to-end on a home
+  broadband line.
+
+The full crawl materialises **roughly 6,700 small JSON / GeoJSON files**
+totaling ~120 MB.
+
+## Stage 2 — parse
+
+```python
+ParseStage(config.parse, raw_dir, parsed_dir).run()
+```
+
+Three jobs:
+
+1. **Normalise Vietnamese-formatted numbers.** The `p.co_dvhc_id` endpoint
+   uses Vietnamese locale (`"6.360,83"` = 6,360.83 km²); the GeoJSON
+   `properties` block uses English (`"575.29"` = 575.29 km²,
+   `"157629"` = 157,629 people). `parse_vi_decimal` and `parse_vi_int` in
+   `packages/scraper/sapnhap.py` cover both idioms.
+2. **Summarise GeoJSON.** Each FeatureCollection collapses to a single row
+   with `centroid_lon`, `centroid_lat`, `bbox`, `geom_type`, `n_vertices`,
+   and (when `parse.flatten_geojson=true`) a shapely-emitted WKT string.
+   We use shapely's true centroid for polygons; for the rare environments
+   without shapely, a ring-walk arithmetic-mean centroid is good enough for
+   plotting.
+3. **Stamp parent-province for every commune & committee.** The
+   `tentinh` attribute on `p.co_dvhc_id` and the `a04_tentinh` attribute on
+   the GeoJSON committee features carry the parent-province name; we
+   resolve it against the 34-row province list to attach a stable
+   `parent_ma` (NSO 2-digit province code).
+
+Output is a single canonical row per entity (province, commune, or
+committee), written as both `parsed.jsonl` and `parsed.parquet`. Schema:
+
+| column          | type     | notes                                                         |
+| --------------- | -------- | ------------------------------------------------------------- |
+| `id`             | str      | feature id (`==` malk)                                        |
+| `kind`           | str      | `province` / `commune` / `committee`                          |
+| `ma`             | str      | NSO 2-digit province code or 5-char commune code              |
+| `ten`            | str      | canonical Vietnamese name                                     |
+| `type`           | str      | `Tỉnh` / `Thành Phố` / `Phường` / `Xã` / `Đặc khu` / …        |
+| `ten_short`      | str      | `ten` with the type prefix stripped                           |
+| `area_km2`       | float    | parsed via `parse_vi_decimal`                                 |
+| `population`     | int      | parsed via `parse_vi_int`                                     |
+| `density`        | float    | `population / area_km2`                                       |
+| `capital`        | str?     | `trungtamhc` (administrative-centre address)                  |
+| `address`        | str?     |                                                                |
+| `phone`          | str?     |                                                                |
+| `decree`         | str?     | `cancu` (e.g. `Nghị quyết số 202/2025/QH15`)                  |
+| `decree_url`     | str?     | usually a `vanban.chinhphu.vn` permalink                      |
+| `predecessors`   | str?     | raw `truocsapnhap` prose                                      |
+| `parent_ma`      | str?     | NSO-code of the parent province (for communes & committees)   |
+| `parent_ten`     | str?     |                                                                |
+| `centroid_lon/lat` | float? | from the geometry summary                                     |
+| `bbox`           | list?    | `[lon_min, lat_min, lon_max, lat_max]`                        |
+| `geom_type`      | str?     | `Polygon` / `MultiPolygon` / `Point`                           |
+| `wkt`            | str?     | shapely WKT (only when `flatten_geojson=true`)                 |
+
+## Stage 3 — extract
+
+```python
+ExtractStage(config.extract, parsed_dir, extracted_dir).run()
+```
+
+Adds the analytical columns the downstream notebook + the visualizer need:
+
+* **`macro_region`** — every entity is mapped to one of the six GSO
+  macro-regions (`northern_midlands`, `red_river_delta`, `central_coast`,
+  `central_highlands`, `southeast`, `mekong_delta`). The mapping table
+  lives in `packages/curator/regions.py` and is hand-curated against the
+  post-merger 34-province list.
+* **`predecessors_list`** — explodes the `truocsapnhap` Vietnamese prose
+  into a deduplicated list of predecessor names. Handles separators
+  (`,`, `và`, `cùng`, `;`), strips mereological qualifiers ("phần còn lại
+  của", "một phần"), and trims trailing "sau khi sắp xếp" clauses.
+* **`n_predecessors`** — `len(predecessors_list)`.
+* **`keywords`** — top-N TF-IDF unigrams + bigrams over the merger-lineage
+  prose; uses a Vietnamese-friendly token pattern
+  (`r"(?u)\b[\wÀ-ỹ]{3,}\b"`) so diacritics survive tokenisation.
+* **`embed_text`** — single canonical Vietnamese descriptor (name + parent
+  + type + predecessors + capital + decree) that the embedding stage
+  consumes.
+
+Output: `extracted.jsonl` + `extracted.parquet`.
+
+## Stage 4 — embed
+
+```python
+EmbedStage(config.embed, extracted_dir, embedded_dir).run()
+```
+
+Encodes every record's `embed_text` field with
+[`sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`](https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2)
+on CPU (384-d, normalised, `batch_size=64`). The full corpus of ~6,700
+short Vietnamese descriptors finishes in ~3 minutes on an M-series Mac.
+
+For NIM-hosted embeddings (e.g. `nvidia/llama-3.2-nv-embedqa-1b-v2`)
+the same backend abstraction used by `personas-vn` would slot in here —
+set `embed.backend: nim` and a `PERSONAS_VN_LLM_API_KEY` env var.
+
+Output: `embedded.parquet` with the meta columns plus a `vector`
+column (list[float]).
+
+## Stage 5 — reduce
+
+```python
+ReduceStage(config.reduce, embedded_dir, reduced_dir).run()
+```
+
+* **UMAP** projection to 2-D with cosine metric (15 neighbours,
+  `min_dist=0.1`, `random_state=20260508`).
+* **Density-based HDBSCAN** clustering (`min_cluster_size=⌊n/80⌋`),
+  emitting an integer `cluster` column with `-1` reserved for low-density
+  noise points.
+
+Output: `reduced.parquet` — every meta column from the embed stage plus
+`x`, `y`, and (when `reduce.cluster=true`) `cluster`. This is the parquet
+that feeds the UMAP plots in [`DATAANALYSIS.ipynb`](DATAANALYSIS.ipynb)
+and the curator-tab in any future Gradio visualizer.
+
+## NeMo Curator backend
+
+Pass `--backend nemo_curator` to `geography-vn curate` and the same five
+stage objects are wrapped as `nemo_curator.core.stage.ProcessingStage`
+sub-classes and handed to a real `nemo_curator.core.pipeline.Pipeline`
+running through `nemo_curator.backends.experimental.in_process.InProcessExecutor`.
+The wire-shape on disk is identical, so the rest of the pipeline (HF
+upload, analysis notebook) does not care which executor ran. To go
+distributed, swap `InProcessExecutor` for
+`XennaExecutor` / `RayDataExecutor` — no code changes needed to the
+stages themselves.
+
+## Re-running individual stages
+
+```bash
+# Re-run only the embed + reduce stages (cheap when the corpus stayed put
+# but the model changed):
+geography-vn curate --only embed reduce
+
+# Re-run everything except the slow geometry crawl:
+geography-vn curate --skip download   # cached anyway, but explicit is faster
+```
+
+The on-disk per-URL cache (`raw/_cache/`) means that even
+`--only download` re-runs are near-instantaneous after the first crawl —
+only newly-published feature ids hit the network.