docs(comparison): add continent-scale Europe section

First continent-scale run lands on the comparison page. Hetzner
bare-metal host (48 cores, 124 GB RAM, NVMe LVM at /mnt/fast),
Geofabrik europe-latest.osm.pbf at 32 GB.

Cairn OSM ingest finishes in 4 h 25 m: pass 0 (max-id scan) → pass
1 (flatnode write) → pass 2a (node-place emit) → pass 2b0
(relation way-ref scan) → pass 2b1 (parallel way-place emit, 22.16 M
ways emitted) → pass 2b2 (admin relations, 363 131 polygons after
dedup, 25.45 M places after dedup). The bundle-write tail (admin
tiles, text index, place tiles, point tiles, manifest, SBOM) is in
flight at commit time; final disk / hot RSS / p95 land in the next
update.

The Phase 6e v2 build pipeline is what makes this run feasible on
the contended cluster — single shared AdminIndex across both
enrichment passes, admin-tile write moved ahead of the tantivy +
place-tile peaks (so the polygon ring set is freed before
allocation pressure rises), and interleaved place-tile + point-tile
emission off one sorted slice (no Vec<PlacePoint> materialization
intermediate). Plus rkyv size_64 to lift the 2 GB rel-pointer
ceiling on Russia / Norway / Greece-scale admin tile archives.

Pelias / Photon / Nominatim Europe measurements: not yet run.
Projected ranges noted in the table (12-18 h, 6-10 h, 24-48 h
respectively) based on incumbent docs and our country-scale
observations. Full parity benches ship in a follow-up.

Caveat block notes the operational knobs needed on this specific
cluster (pause airflow Python workers during the build window;
bump vm.max_map_count to 1 048 576 so Cairn's mmap pressure
doesn't run into the 65 530 default). They're host-config, not
algorithmic, and don't show up on country-scale runs.

Author: marcosci

comparison/index.html

@@ -454,6 +454,85 @@ <h3>Recall on noisy queries</h3>
         gotchas is at
         <a href="https://github.com/cairn-geocoder/cairn/blob/main/benchmarks/results/REPORT.md"><code>benchmarks/results/REPORT.md</code></a>.
       </p>
+
+      <h3>Continent scale — Europe</h3>
+      <p>
+        First continent-scale run, on a different host: a Hetzner
+        bare-metal box with 48 cores, 124 GB RAM, 32 GB swap (rotating)
+        + 64 GB NVMe swap, NVMe LVM volume mounted at
+        <code>/mnt/fast</code>. Input is Geofabrik
+        <code>europe-latest.osm.pbf</code> (32 GB), 7× larger than
+        Germany. Cairn runs against this PBF through the same
+        <code>cairn-build</code> binary as the country-scale runs,
+        with rkyv <code>size_64</code> rel-pointers (lifts the per-tile
+        2 GB ceiling that 32-bit relative pointers impose on Russia /
+        Norway-scale admin polygons) and the Phase 6e streaming
+        pipeline (single shared AdminIndex, admin-tile write before
+        tantivy + place-tile write, interleaved place-tile + point-tile
+        emission off the same sorted slice).
+      </p>
+      <table class="api">
+        <thead>
+          <tr>
+            <th>Engine</th>
+            <th>OSM ingest</th>
+            <th>Places after dedup</th>
+            <th>Admin polygons</th>
+            <th>Bundle disk</th>
+            <th>Hot RSS</th>
+            <th>p95</th>
+          </tr>
+        </thead>
+        <tbody>
+          <tr>
+            <td><strong>Cairn</strong></td>
+            <td><strong>4 h 25 m</strong></td>
+            <td><strong>25.45 M</strong></td>
+            <td><strong>363 131</strong></td>
+            <td><em>pending — build in flight</em></td>
+            <td><em>pending</em></td>
+            <td><em>pending</em></td>
+          </tr>
+          <tr>
+            <td>Pelias</td>
+            <td colspan="6"><em>not run yet — projected ~12–18 h on the same hardware (Elasticsearch import + WhosOnFirst + OpenAddresses + OSM importers run sequentially via docker-compose; 32 GB PBF dominates the OSM importer pass)</em></td>
+          </tr>
+          <tr>
+            <td>Photon</td>
+            <td colspan="6"><em>not run yet — projected ~6–10 h (Nominatim DB import is the long pole; Photon's Elasticsearch index build is fast once the Postgres source exists)</em></td>
+          </tr>
+          <tr>
+            <td>Nominatim</td>
+            <td colspan="6"><em>not run yet — projected ~24–48 h on a single Postgres instance (osm2pgsql import + indexes; cluster has the I/O headroom but the runtime is single-process bound)</em></td>
+          </tr>
+        </tbody>
+      </table>
+      <p>
+        Cairn's OSM ingest on Europe — pass 0 (max-id scan) → pass 1
+        (flatnode write, 110 GB virtual / ~88 GB on NVMe sparse) →
+        pass 2a (node-place emit) → pass 2b0 (relation way-ref scan)
+        → pass 2b1 (parallel way-place emit, 458 M ways scanned,
+        22.16 M emitted) → pass 2b2 (admin relations, 363 K polygons
+        out of 8.31 M relations) — completes in 4 h 25 m on this box,
+        sustaining ~2300 % CPU (≈ 23 cores) through the
+        rayon-parallel passes. Final bundle write (admin tiles, text
+        index, place tiles, point tiles, manifest, SBOM) runs after
+        OSM ingest finishes; full numbers + the corresponding
+        Pelias / Photon / Nominatim measurements ship in the next
+        update.
+      </p>
+      <p class="caveat">
+        The cluster also runs an Airflow-driven planetary-imagery
+        pipeline on the same host. We pause those workers
+        (<code>kill -KILL</code> the Python pool, airflow respawns
+        them after the build window) and bump
+        <code>vm.max_map_count</code> to 1 048 576 so Cairn's mmap
+        pressure (110 GB flatnode + per-tile output mappings + tantivy
+        segment files) doesn't run into the 65 530 default limit.
+        Both knobs are operational, not algorithmic — they don't show
+        up in country-scale runs because the cluster has more
+        headroom there.
+      </p>
     </div>
   </section>