de:wasserversorger-discover

Population coverage rose from 87.81% to 88.42% (+0.61pp) after a phase-2b batch of 60 Gemeindeverband units, claiming 115 more municipalities (residual now 2,222 municipalities / 9.68M people). Two synchronous subagents did the research; the mandatory pre-merge cross-shard collision check caught and fixed 3 duplicate company mints (Lechfeld, oberer Neckar, Kleine Kinzig — all remapped to their pre-existing canonical IDs) that the shards' own dedup grep had missed, leaving exactly the 17 expected domain collisions (16 historical + 1 new legitimate shared-VG-portal case). Journal updated with next-queue state, a newly flagged Iller-Rißtal duplicate for human review, and the still-recommended full-roster consolidation pass.

Scheduled run. Read CLAUDE.md, this journal (via tail — full file exceeds the 256KB read cap), and SKILL.md first. Confirmed on-disk state before starting: 2011 companies, 87.81% population coverage (from the 20:30 enrich run, itself on top of the 18:00 vg8 discover run). Re-ran coverage.mjs + build_gv_layer.mjs fresh to confirm queue state matched the journal's prediction (Kellinghusen settled-negative still top, then Untersteinach/Trave-Land/Eidertal carried over unresolved, then VVG der Stadt Uhingen) — matched exactly.

Batch build: build_batch_input.mjs 61 vg9, dropped Kellinghusen (still the settled negative) from the top, leaving 60 units split 30/30. Top of queue was mostly small 2-member Baden-Württemberg VVG/GVV units plus a Bavaria/Sachsen/Thüringen singleton-pair tail — the phase-2b "hard residual" regime, much lower yield per unit than the earlier BW VVG batches (most units here are 2-member).

Split into two synchronous subagents (run_in_background: false on both, dispatched in one message, blocked on both before continuing), 30 units each, with the mandatory no-Agent-tool line in both prompts and disjoint company_id ranges (A: wv-de-10001–10099, B: wv-de-10501–10599). Neither shard recursed into sub-subagents; both ran long (17–24 min turns, ~1030s and ~1425s duration).

Shard results: Shard A resolved 59/65 link rows (63 seeded municipalities + 2 legitimate dual-supplier extra rows), 6 honestly unresolved (the carried-over Trave-Land/Eidertal SH village trio pairs — tried a fresh angle per the brief, genuinely nothing published). Resolved the Untersteinach (VGem) Bavarian trio (Guttenberg/Kupferberg/Ludwigschorgast) this time via each town's own Wasserabgabesatzung (medium confidence, no pdftotext/python3 in this container to extract the PDFs cleanly) — self-run Eigenbetrieb each, not a "Pressecker Gruppe" as a prior run's disproven lead suggested. Shard B resolved 68/71 link rows, 3 honestly unresolved (Sachsen Erzgebirge villages). Correctly applied the Betriebsführung-vs-Konzession distinction again (Wyhl/badenovaNETZE, Allershausen/WZV Paunzhausen technical-only). Flagged a likely pre-existing duplicate for a human: wv-de-1743/wv-de-7036 (Zweckverband (Trinkwasserversorgung/Wasserversorgung) Iller-Rißtal, same seat address) — not fixed (out of this batch's scope, both already on disk from prior runs, not something either shard minted).

Self-caught + orchestrator-caught duplicate mints, fixed pre-merge: shard A avoided 9 and shard B avoided 11 likely re-mints by grepping roster.jsonl first (listed in shard reports). Despite that, the mandatory cross-shard + against-existing-roster collision check (run before consolidate_roster.mjs, per the skill's standing instruction) still caught 3 new duplicate mints that both shards' own dedup passes missed because the existing rows used a differently worded name for the identical entity:

• wv-de-10011 ("Wasser und Abwasserzweckverband Lechfeld") = pre-existing wv-de-7604 ("Wasserzweckverband Lechfeld"), same domain wasser-abwasser.lechfeld.de, same ~20k-resident member list.
• wv-de-10012 ("Zweckverband Wasserversorgung am oberen Neckar") = pre-existing wv-de-8528, identical name, same domain, near-identical member list (Dietingen etc).
• wv-de-10509 ("Zweckverband Wasserversorgung Kleine Kinzig") = pre-existing wv-de-1739 ("Wasserversorgung Kleine Kinzig (Zweckverband)"), same domain zvwkk.de, same 32-member/ 350,000-resident bulk-supply description.

Fixed by remapping each duplicate's link rows (7 total) to the canonical existing company_id and dropping the 3 duplicate roster rows, all before running map_roster.mjs/consolidate_roster.mjs. Re-ran the domain-collision check after the fix: 17 total (16 historical carried unchanged + 1 new legitimate case, vg-wackersdorf.de shared by two distinct self-supplying Gemeinden — Steinberg am See and Wackersdorf — on one shared VG portal, same shape as the existing oberbergkirchen.de/vg-schoensee.de cases). Zero new unexplained collisions.

Merge: map_roster.mjs: 10237 total link rows (was 10101), 232 needing review (228 unmatched — mostly newly-harvested Ortsteil/no-AGS names from this batch's member lists, not a regression — 3 expanded_qualifier, 1 ambiguous: Breitenbrunn, still unresolved, still needs a human pick). consolidate_roster.mjs: 2190 shard rows (105 shard-level dupes, last-shard-wins) → 2085 distinct companies (+74, consistent with 65 net new mints from this batch plus prior-run shard rows that hadn't previously been counted as distinct — see the 1[01]xxx-range grep, 74 IDs in that range). coverage.mjs: population 87.81% → 88.42% (+0.61pp, smallest single-batch gain of the phase-2b run so far — expected, this batch was mostly 2-member units with residual populations in the 7-20k range, not the earlier large multi-member/singleton batches). Municipalities claimed 8606 → 8721 (+115). Residual: 2337 → 2222 municipalities, 10.192M → 9.678M people — first time under 2,250 residual municipalities and under 9.7M residual population. pending_domain in the roster: 158 (up from the ~54 reported after the last enrich run, since this batch minted several new pending_domain rows — enrich hasn't run against this batch's output yet). build_gv_layer.mjs (fresh for next run): units with residual members 2103 → 2047, multi-member 229 → 171 (covering 346 municipalities, was 463), singletons 1874 → 1876.

For next run: New queue top: Kellinghusen (still the settled negative, skip it), Trave-Land (3, still unresolved — SH villages, two consecutive runs found nothing published), Eidertal (3, same), then a run of 2-member GVV/VVG units (GVV Rot-Tannheim, VVG der Stadt Zell im Wiesental, etc) — the multi-member layer is thinning fast (463 → 346 municipalities in one batch) and phase 2b is increasingly singleton-shaped, consistent with the skill's prediction. company_id max on disk now 10540 (sparse: 10001–10039 used by shard A minus 2 dedup drops, 10501–10540 by shard B minus 1 dedup drop) — next batch should mint from 11001+ to stay clear of every prior range; check roster.jsonl before minting regardless, as always. Open duplicate-mint/collision list: the 2 long-standing human-review items (wv-de-1415/7505 Wiesenbachgruppe, wv-de-3925/8502 Schwabachgruppe) plus the wv-de-1148/8515 Bad Königshofen pair from the 20:30 run, plus now a new flagged pair for review: wv-de-1743/7036 (Iller-Rißtal, near-identical Zweckverband names, same seat address, found by shard B's own dedup grep but not mergeable without a human call on which name is authoritative). The still-overdue human consolidation pass (scripted domain-collision + near-identical-name sweep across the full 2085-row roster) remains recommended and keeps growing more valuable — this run is the second in a row where the automated per-batch collision check caught mints that both shards' own dedup grep missed, which is exactly the failure mode that pass would eliminate structurally instead of per-batch.