```⏺ Complete Summary

  Project Overview

  Traccar Geocoder — a self-hosted reverse geocoding service. C++ builder parses OpenStreetMap PBF files into binary index files, Rust server
  mmap's them for sub-millisecond queries.

  What We Built (in order)

  1. --admin-only and --no-addresses flags — Builder can produce stripped-down indexes (admin boundaries only, or streets without house numbers).
   Server auto-detects which files are present.
  2. Multi-output builder rewrite — Single PBF parse produces all 3 index modes (full, no-addresses, admin-only) + per-continent subsets. Binary
  cache (--save-cache/--load-cache) avoids re-parsing. --multi-output and --continents flags.
  3. Parallelism improvements (iterative):
    - Parallel deduplication of cell maps (std::async)
    - Parallel index writing — 3 modes written simultaneously
    - Parallel continent filtering — 8 continents processed simultaneously
    - Concurrent admin polygon S2 covering — thread pool processes S2 coverings while PBF reader continues
    - Streaming parallel node processing — producer-consumer queue, workers process PBF blocks in parallel
    - Streaming parallel way processing — same pattern for ways
    - Dense array node index — 100GB+ mmap with MAP_NORESERVE, lockless writes (replaced FlexMem which OOM'd, and sharded FlexMem which had too
  much overhead)
    - Parallel admin polygon assembly — custom ring stitcher replaces osmium's sequential multipolygon assembler, processes 778K relations across
   all cores
  4. Node ID matching for ring stitching — Fixed coordinate-based matching (lossy floats) to use OSM node IDs (exact integers). Stores first/last
   node IDs in way geometry data. Only stores way geometries for admin boundary member ways (~3M vs ~300M total ways).
  5. Ring rotation normalization — add_admin_polygon normalizes ring starting vertex before Douglas-Peucker simplification for consistent output.
  6. Live dashboard — Python HTTP server at localhost:8888 with CPU/memory graphs, elapsed timers (tick between refreshes), log output per node.
  Polls 4 SSH nodes every 30s.
  7. Semantic comparison tool (compare_indexes.cpp) — Decodes binary index files, sorts by canonical keys, compares vertex rings as sets (handles
   rotation/direction), tolerant float comparison.
  8. Kubernetes manifests (k8s/planet-server.yaml) — PVC-backed deployment that downloads planet PBF once and serves via HTTP for other pods.

  Current State of Nodes

  - Node 1 (mich.immich.cloud:30022, 24c/125GB): Running the original sequential code, ~15+ hours in, at 68.7% PBF. SSH: ssh -o
  ControlPath=/tmp/ssh-control/mich -p 30022 mich.immich.cloud
  - Node 2a (10.10.10.12:30022, 48c/192GB cgroup): Just reconnected, fresh pod. SSH: ssh -o ControlPath=/tmp/ssh-control/newbox -p 30022
  root@10.10.10.12. OIDC auth pending (code GKCS-KBJW).
  - Node 2b (10.10.10.11:30022, 48c/192GB cgroup): Available but not connected yet.
  - Node 3 (100.113.88.121, password elon-is-insane, user michtest, 32c/174GB, no cgroup, 50GB swap): SSH: ssh -o
  ControlPath=/tmp/ssh-control/node3 michtest@100.113.88.121. Just finished parallel Germany build. Sequential Germany build should be started.
  Has saved outputs from the old planet run at ~/indexes-old-code/ and ~/planet-old-code.cache.

  The Current Problem

  The parallel admin assembly produces ~12K fewer admin polygons than osmium's sequential assembler on planet data (931,963 vs 943,638). Analysis
   shows:
  - 92K "missing" (in old, not in new)
  - 80K "extra" (in new, not in old)
  - Net: ~12K fewer
  - Spread across all admin levels
  - Top missing names are small plot boundaries

  The node-ID matching fix was just deployed but hasn't been tested on a large dataset yet. The Germany build on node 3 just completed with the
  new code — needs comparison against the sequential version.

  Files on Node 3