de:wasserversorger-discover

Discovery run complete: worked the top 60 phase-2b Gemeindeverband queue units (all singleton municipalities, sorted by residual population), dispatched two synchronous shards, and caught 4 duplicate company mints plus 13 self-caught duplicates via the mandatory domain-collision check before merging. Population coverage rose 89.70% → 90.57% (crossed 90%), municipalities claimed 9074 → 9136, roster grew 2284 → 2326 companies, with 0 new domain collisions after fixes. Journal updated with the new host traps (hirschaid.de, wlv-helmsdorf.de, rheinau.de), a CSV-quoting gotcha in queue sorting, and next-run guidance (Bayern/Thüringen/Hessen remain the biggest levers).

Triggered via chat (users/108934946863892655718), not the schedule. Read CLAUDE.md and the tail of this journal, then SKILL.md in full. On-disk state before starting: 2284 companies, 89.70% population coverage, 9074/10943 municipalities claimed. Phase 2a (singleton units, residual population ≥10,000) is now essentially exhausted — only 1 unit (Bannewitz, 11,122) still qualified, so this run is the first squarely in phase 2b: build_gv_layer.mjs showed 11 multi-member units left covering just 25 municipalities, everything else is singleton.

Batch selection caveat worth flagging for future runs: a first attempt at picking the top-60 queue rows with plain sort -t',' -k6 -n -r on gemeindeverband-queue.csv gave a WRONG order — several unit names (GVV Hardheim-Walldürn, Neustadt in Sachsen, Stadt, etc.) contain commas inside quoted CSV fields, which shell sort -t',' doesn't respect, silently shifting the population column for those rows. Wrote a tiny .mjs (build_batch60.mjs, kept in the campaign dir) using lib.mjs's real CSV parser instead, cross-referenced against current claims (links.jsonl/roster-links.jsonl/longtail.jsonl) to get the correct top-60 by residual_population, with each unit's still-unclaimed member AGS resolved from gemeindeverband-members.csv. All 60 selected units turned out to have exactly 1 residual member each — phase 2b at this population tier is pure singleton work now, no more multi-member yield.

Dispatch: two synchronous subagents (run_in_background: false, one message, blocked on both), disjoint company_id ranges (A: 14001–14299, B: 14501–14799), mandatory no-Agent-tool line in both prompts (_tasks/netzbetreiber-contact/german-water/ still absent from this checkout — inlined the full method again). Neither shard recursed. Shard A ran ~13 min (146 tool uses), shard B ran ~11 min (127 tool uses).

Shard results: Shard A resolved 30/30 fully, 29 new companies (wv-de-14001–14029), 1 reuse (Bad Fallingbostel → pre-existing wv-de-3285, self-caught). New host traps: hirschaid.de 403s, wlv-helmsdorf.de 500s on every attempt. Shard B resolved 30/30 fully, but its own dedup-grep ran only after initial minting rather than before — it self-caught 13 of its own 30 candidate mints as pre-existing companies (Hagenow, Ludwigslust-Grabow, Orla, Anklam, TAVOB, Pritzwalk, Pirna/Sebnitz, VersorgungsBetriebe Elbe, Hildburghausen, Bad Dürrenberg, Rheingauwasser, Glückstadt, Wittenhorst) and rewrote its output to reuse the existing IDs before handing off. New host trap: rheinau.de 403s.

Mandatory domain-collision check, run before consolidate_roster.mjs, across the full existing roster + both new shard files, found the exact failure mode this skill's banner warns about: shard A's own dedup pass (checked only against roster.jsonl, not thoroughly enough / or missed due to a formatting difference) still let through 4 duplicate mints that collided with pre-existing companies: wv-de-14011 (Böhmfelder Gruppe, dupe of wv-de-11532), wv-de-14021 (WWAZ, dupe of wv-de-1201), wv-de-14025 (Stadtwerke Vilsbiburg, dupe of wv-de-5503), wv-de-14027 (ZWAS, dupe of wv-de-636) — all four verified as genuine same-entity/same-domain matches before fixing. Fixed by hand: removed the 4 duplicate rows from roster.part-a.jsonl and remapped their 13 associated link rows (including bonus-harvest members) to the pre-existing company_ids. Re-ran the collision check after the fix: 25 total, all pre-existing and previously reviewed as benign (shared VG/Amt portal or open review-queue items) — 0 new collisions once the shard-A duplicates were corrected. This is the second run in a row where the mandatory re-check (not trusting a shard's self-report) caught something a shard missed — the rule earns its keep every time it's applied.

Merge: map_roster.mjs: 11064 total link rows (was 10892), 323 needing review (mostly newly-harvested no-AGS bonus names, not a regression); 6 ambiguous entries found on inspection, all pre-existing from 2026-08-26 through 2026-08-28 runs (Breitenbrunn, Steinbach, Hausen, Mühlhausen, Weingarten, Papendorf) — none from this run's new companies, just previously-unflagged review-queue growth the printed summary truncates at 25 rows (had to grep roster-links.jsonl directly for match_method":"ambiguous" to see them all). consolidate_roster.mjs: 2534 shard rows (208 shard-level dupes) → 2326 distinct companies (+42 from 2284). coverage.mjs: population 89.70% → 90.57% (+0.87pp — crossed the 90% line). Municipalities claimed 9074 → 9136 (+62, more than the 60 dispatched thanks to bonus member-list harvests from Böhmfeld, WWAZ, and others). Residual: 1869 → 1807 municipalities, 8.611M → 7.878M people. build_gv_layer.mjs: multi-member units unchanged at 11/25 (this batch was singleton work, as expected), singletons 1844 → 1782. By-Land, lowest population coverage: Bayern 76.2% (1326/2217), Thüringen 79.6%, Hessen 80.8%, Sachsen-Anhalt 81.1% — Bayern remains the biggest lever by a wide margin, unchanged for many runs.

For next run: phase 2b singleton work continues — next top-60 by residual_population should be re-derived with build_batch60.mjs (already in the campaign dir, do not shell-sort the queue CSV directly — the comma-in-quoted-field trap above). Largest still-unclaimed municipalities after this run: Gräfenhainichen (11,228), Wittingen (11,136), Coswig (Anhalt) (11,129), Bannewitz (11,122), Bad Belzig (11,118), Schöningen (11,078), Werther (Westf.) (11,051), Boizenburg/Elbe (11,048) — all comfortably in reach of another 60-unit batch. New host traps to add to the known list: hirschaid.de (403), wlv-helmsdorf.de (500), rheinau.de (403). Open duplicate-mint/collision list unchanged from prior runs (six entries, see prior journal entries) — the overdue human consolidation pass remains the single biggest lever for cleaning up the accumulated review queue.