de:wasserversorger-enrich

Journal entry written. The run is complete: all 45 companies enriched, verified, merged, and the campaign pipeline re-run clean.

Enriched 45 German water suppliers this pass (0 pending_domain, drawn purely by population priority): enrichment.jsonl grew to 1709 records, roster-links.jsonl to 19402, population coverage holds at 99.99% (unchanged as expected — this batch enriched already-claimed companies rather than expanding the roster). All 8 research sub-agents completed cleanly with zero non-suppliers found and several corrections to stale prior-pass leads (notably Eppstein's Handelsregister claim and the Pöringer Gruppe member-count). One process hiccup: my two top-level research agents each silently fanned out into 4 nested background sub-agents instead of researching directly — caught via ListAgents and resolved by blocking on all 8 child tasks directly, so no work was lost, but it's now logged as a recurring pattern to watch for in future runs.

build_enrich_batch_excl.mjs 45 → 0 pending_domain (queue stayed empty since 2026-09-03 20:30), batch drawn purely by population priority (13.0k-13.2k _pop_served band). Split into shard A (23) / shard B (22), each dispatched to one top-level Agent call with run_in_background: false as the 2026-09-03 lesson prescribes.

Process note, same trap as 2026-09-03 recurring in a new shape: both top-level agents independently decided to fan out their batch into 4 nested sub-agents each (8 total) rather than researching directly, and their first-round reports described this as already done ("I've launched N parallel research agents... I'll report back once they finish") — i.e. they tried to end their own turn while children were still running, the exact anti-pattern SKILL.md's Resumability section warns the top-level run about, just one level down. run_in_background: false on my Agent call did not prevent the child agent from spawning its own background grandchildren. Caught it immediately via ListAgents (8 running subagents where 2 were expected) and blocked on all 8 task_ids directly with TaskOutput(block:true, timeout:600000) in parallel until every one reported completed. Nothing was lost, but it cost an extra round-trip. Lesson for future runs: after any Agent call returns, treat a suspiciously fast or delegation-shaped response as a signal to check ListAgents before trusting "done" — this can happen at any nesting depth, not just the top level.

Verification before merge: enrichment.jsonl 1664 → 1709 lines, all 45 batch company_ids confirmed present with zero JSON corruption and zero collisions within the batch (3 unrelated pre-existing dup ids — wv-de-7559 ×3, wv-de-7554 ×3, wv-de-2600 ×2 — predate this run, normal append-only re-research artifacts, none touched this pass). 5 roster-links part shards (224 lines total: 103 shard A + 41+37+21+22 shard B sub-shards) all validated as clean JSON before map_roster.mjs merged them.

Merge: map_roster.mjs: 19178 → 19402 roster-links rows (+224). consolidate_roster.mjs: still 177 shards / 3375 distinct companies (unchanged — expected, this batch enriches already-claimed companies, adds no new roster rows); roster-links.csv 16214 → 16395 distinct after dedup (+181 net). coverage.mjs: population 99.99% → 99.99% (unchanged, expected), 24 municipalities / 11,023 people still residual, same list as last run (Hartmannsdorf-Reichenau largest at 1,040).

Non-suppliers found this batch: 0 (contrast with 2 in the prior 2026-09-03 20:30 batch) — all 45 confirmed genuine drinking-water operators.

Field recovery this batch (45 companies, 8-subagent breakdown): name/address/phone ~45/45; email ~38/45 (several JS/image/TYPO3-token-obfuscated, decoded where feasible — Dossenheim's required cross-verifying a Caesar-2 letter-shift cipher against two independently-plaintext tokens on the same page); emergency_phone_number ~40/45, multiple only found after following a separately-linked Notdienst/Bereitschaftsdienst/Notruf page rather than the main water page (the known false-negative trap held again — Hamm-Sieg, Manching, Zehdenick, Winnweiler all needed it); contact_form_url ~12/45 real forms; analysis_url found for ~40/45 with real evidence; multi_zone true for ~18/45.

Corrections vs. carried-forward leads (batch input included prior-pass leads as unverified navigation hints, not facts — treated accordingly):

• wv-de-3981 Altötting: operator's own current page states the joint "Gemeinschaftliches Wasserwerk Altötting-Neuötting-Winhöring" supplies all three municipalities' Ortsnetz directly, contradicting the prior lead's "administratively separate billing" assumption.
• wv-de-3986 Eppstein: prior pass's handelsregister_nr: not_applicable + legal_form_verbatim: "Körperschaft des öffentlichen Rechts" did not hold up — current Impressum has neither phrase nor HR number; corrected to not_found.
• wv-de-3993 OeWA Oebisfelde: real Versorgungsgebiet (~6,800 residents, former Stadt Oebisfelde only) is much smaller than the batch's _pop_served (13,002, reflecting the full merged Einheitsgemeinde Oebisfelde-Weferlingen) — a scope mismatch worth a maintainer's attention if _pop_served feeds any downstream metric.
• wv-de-7716 Pöringer Gruppe: prior claim of "16 further municipalities/hamlets" not corroborated by either the operator's own site or vg-puergen.de — both independently list exactly 4 Mitgliedsgemeinden; the extra names were Ortsteile/sampling points, not members.
• wv-de-412 Winnweiler: water-supply Betrieb transferred to AöR "BWGEW" (Göllheim/Eisenberg/Winnweiler) as of 2023-01-01 — more specific than the prior lead.
• wv-de-13524 Sulz am Neckar: several "Stadtteile" are bulk-supplied by a separate Zweckverband Kleiner Heuberg, not Sulz's own wells — flagged, not silently merged.
• wv-de-3650 Kerken: claimed Stadtwerke Kamp-Lintfort operational takeover was not found on any Gemeinde Kerken page this pass — left unconfirmed rather than carried forward.
• wv-de-4005 Neunkirchen (Siegerland): WVS wholesale-supply hint confirmed (Wasserverband Siegen-Wittgenstein bulk-supplies to the Gemeindegrenze, Neunkirchen retails). Also: wvs.nrw rendered fine via plain curl this pass, contradicting an older journal note calling it an inaccessible JS SPA — site behavior may have changed, or the earlier note was itself wrong.
• wv-de-14002 Hettstedt: confirmed 100% bulk water from Fernwasserversorgung Elbaue-Ostharz GmbH (FEO, Wasserwerk Wienrode); the FEO-hosted analysis PDF is an unreadable image scan — noted, not guessed.

New host/encoding traps, all measured this pass:

• stadtwerke-winterberg.de HTTPS still times out (curl exit 28, reproducing an old trap) but HTTP 301-redirects cleanly to rathaus-winterberg.de, which holds all real content — first working resolution for this host.
• wasserportal.info (BDEW's shared Angular SPA, used by Bebra, Rotenburg a.d. Fulda, and Lauterbach for third-party analysis hosting) renders nothing via curl/WebFetch — no headless browser available in this toolset, so those three got lower-confidence analysis records rather than fabricated values.
• neunkirchen-siegerland.de serves ISO-8859-15 with high-byte content that makes plain grep silently treat the file as binary — needs grep -a.
• dossenheim.de (TYPO3/Hirsch & Wölfl) cloaks all mailto links via a token cipher rather than plain mailto: hrefs — decoded and cross-verified against two independently-plaintext tokens on the same page.
• Dillenberggruppe's download.php PDF endpoint returned HTTP 403 to both WebFetch and curl (with/without Referer) — not a captcha, not bypassed, just inaccessible.
• sb-roding.de (wv-de-3977) is a full JS/SPA shell (every path returns an identical loading page) and stadt-badlaasphe.de (wv-de-3979) returns a hard 403 bot-challenge on every path — neither bypassed; both corroborated via third-party listings with confidence downgraded.

Still stale: export_csv.mjs/map_ags.mjs still point at the dead legacy companies.jsonl/links.jsonl pair, not roster.jsonl/roster-links.jsonl — this run used the live map_roster.mjs → consolidate_roster.mjs → coverage.mjs pipeline instead, as every recent run has done (11+ consecutive runs now). Still worth a maintainer fixing.

For next run: continue with build_enrich_batch_excl.mjs 45. 0 pending_domain. New human-queue items from this pass: OeWA _pop_served scope mismatch, Sulz-am-Neckar bulk-supply Stadtteile, Kerken Kamp-Lintfort claim unconfirmed. Standing items from prior runs (Beelitz/ Meerane roster defects, Solms durable host issue, Linden/Dahner Felsenland CMS migrations, Glückstadt Wasserwerk disagreement, WWS Saarwellingen entity-naming ambiguity, Gauting/Würmtal re-key, Gilching duplicate, Heilbad Heiligenstadt non-supplier suspicion, wv-de-16016 Ainring / wv-de-3724 Prenzlau self_only contradictions, Oberasbach RangauWerke freshness gap, Sinzig/ Rhein-Ahr merger, Kronach HRB conflict, Hille/Hartum mis-key, Woringen Ungerhausen contradiction, Halver/Kierspe naming coincidence, Römerberg-Dudenhofen domain collision, Grefrath 403 trap, Raunheim CMS migration, WVS Siegen-Wittgenstein own-enrichment gap, Krumbach/Stauden double-claim, Neutraubling self-supply unconfirmed, vgwerke-weilerbach.de trap) still open, none touched this pass.