de:wasserversorger-discover

Run complete. Population coverage rose 92.18% → 92.83% (+0.65pp), municipalities claimed 9,282 → 9,342 (+60), and the roster grew to 2,449 distinct companies (+40) after fixing 5 duplicate mints caught by the mandatory post-merge collision check (0 erroneous collisions remained after fixing). Bayern remains the biggest residual lever (79.6% covered); a journal entry with full details, new host traps, and next-batch pointers has been appended for the next run.

Triggered via chat, not the schedule. Read CLAUDE.md and the tail of this journal, then SKILL.md in full. On-disk state before starting matched the 18:00 run's reported final numbers exactly (2409 companies, 92.18%, 9282/10943, coverage.json mtime ~20:19 CEST — consistent with UTC-labeled run IDs: this run's own "18-39-31" is UTC, actual wall clock was 20:39 CEST). Fresh start, nothing unmerged. build_gv_layer.mjs: still only 11 multi-member units / 25 municipalities — phase 2b top tier remains pure singleton work.

Batch selection: build_batch60.mjs gave the next top-60 by residual_population: Mücke (9,267) down to Wenzenbach (~8,7xx). All 60 singleton (1 residual member each) — top tier continues thinning below 9,300/unit.

Dispatch: two synchronous subagents (run_in_background: false, one message, blocked on both), fresh slice names s17a/s17b, company_id ranges A: 17001–17299, B: 17501–17799 (previous max on disk was 16520). Mandatory no-Agent-tool line inlined in both prompts. Neither shard recursed. Shard A ran ~14.6 min (142 tool uses), shard B ran ~20.9 min (203 tool uses).

Shard results: Shard A resolved all 30, minted 20 companies initially (wv-de-17001–17020), 10 genuine self_supplied (Mücke, Lügde, Hilzingen, Lichtenstein, Altenbeken, Freystadt core, Nittenau, Rain, Florstadt, Zwiesel, Rehau, Hohenbrunn — some overlap in that list per the subagent's own slightly-miscounted summary, but the roster file is authoritative), plus a genuine Salzhemmendorf split (2 suppliers by district). New host traps: two same-name "Lichtenstein" municipalities (Sachsen vs BW) needing Kreis-qualified search; florstadt.de JS-heavy defeating direct fetch; fuewasser.de name-guess trap (wrong region, Fürstenwalde not Barleben). Shard B resolved all 30, minted 25 companies initially (wv-de-17501–17525+), 12 self_supplied, 8 reused existing company_ids correctly (no new mint). New host traps: two same-name "Neuried" municipalities (Bayern vs Ortenaukreis BW); wasser-freiberg.de 403s to WebFetch (bot protection); borgholzhausen.de site restructure broke old deep links (Nov 2025); liederbach.eu 403s WebFetch but curl works (same pattern as the documented pluederhausen.de trap).

Mandatory domain-collision check, run before consolidate_roster.mjs, across the full existing roster + both new shard files — again caught what the shards' own self-checks missed, fifth run in a row this has happened: 5 of this run's "new" mints were genuine duplicates of pre-existing companies, all same-domain/same-name/same-entity on inspection: wv-de-17013 (Stadt Rain Wasserwerk Regiebetrieb, dupe of wv-de-8536 "Stadt Rain — Wasserwerk (bulk supplier to Niederschönenfeld)" — same physical Wasserwerk serving two roles, retail to Rain and wholesale to Niederschönenfeld; the pre-existing row already models this as one company, shard A should have reused it), wv-de-17521 (Wasserzweckverband Freiberg, dupe of wv-de-3176), wv-de-17512 (Wasser- und Abwasserzweckverband Saalkreis, dupe of wv-de-3666), wv-de-17524 (Zweckverband zur Wasserversorgung Landkreis Regensburg-Süd, dupe of wv-de-5505), wv-de-17508 (Zweckverband Wasserversorgung und Abwasserbehandlung Rügen/ZWAR, dupe of wv-de-819). One additional same-domain hit (entega.ag: pre-existing wv-de-3014 "ENTEGA AG" vs shard A's new wv-de-17017 "ENTEGA Wasserversorgung Biblis GmbH") was verified as a legitimate parent/subsidiary pair — distinct HRB-registered legal entity dedicated to Biblis, using the parent's domain for its own published Preisblätter — and correctly kept as two companies, not merged. Fixed by hand: removed the 5 duplicate roster rows (shard A 20→19, shard B 25→21) and remapped their 5 link rows to the pre-existing company_ids. Re-ran the collision check after the fix: 26 total (was 25 pre-existing/benign + the 1 confirmed-legitimate entega.ag pair), 0 new erroneous collisions. Take for future runs, now confirmed a fifth time: a shard's own "grep roster.jsonl before minting" step is not reliable enough on its own — do not skip the mandatory post-merge re-check or trust a shard's "0 collisions" self-report. Also note the Rain case as a new duplicate shape: the same physical entity correctly modeled as one company playing two roles (retail + wholesale) in the pre-existing data, re-minted by a shard that only saw the retail side — when a singleton is a known bulk supplier to a neighbor (as flagged in wv-de-8536's own notes), check the existing roster before minting even for a "new" name-looking match.

Merge: map_roster.mjs: 11335 total link rows (was 11273), 352 needing review (mostly newly-harvested no-AGS names, not a regression), 7 ambiguous entries, all pre-existing (none new this run). consolidate_roster.mjs: 2657 shard rows (208 shard-level dupes) → 2449 distinct companies (+40 from 2409). coverage.mjs: population 92.18% → 92.83% (+0.65pp). Municipalities claimed 9282 → 9342 (+60, exactly the batch size). Residual: 1661 → 1601 municipalities, 6.537M → 5.996M people. build_gv_layer.mjs: multi-member units unchanged at 11/25 (singleton work again), singletons 1636 → 1576. By-Land, lowest population coverage: Bayern 79.6% (1374/2217), Thüringen 81.9%, Hessen 85.4%, Sachsen 87.0%, Sachsen-Anhalt 91.0% — Bayern remains the biggest lever by a wide margin, unchanged for many runs.

For next run: phase 2b singleton work continues, next top-60 should be re-derived with build_batch60.mjs fresh (do not shell-sort the CSV). Largest still-unclaimed after this run, all now below 8,750: Rimbach 8,738, Langweid a.Lech 8,737, Herzberg (Elster) 8,732, Eslohe (Sauerland) 8,717, Großenlüder 8,714, Heiden 8,704, Raguhn-Jeßnitz 8,702, Keltern 8,693 — per-batch population yield keeps shrinking as expected. company_id max on disk now 17525 (sparse: 17001–17020 minus 1 removed in shard A, 17501–17525 minus 4 removed in shard B) — next batch should mint from 18001+. New host traps to add to the known list: two same-name-collision pairs (Lichtenstein Sachsen vs BW; Neuried Bayern vs Ortenaukreis BW — both need Kreis-qualified search), florstadt.de (JS-heavy), fuewasser.de (name-guess trap, wrong region), wasser-freiberg.de (403 to WebFetch, bot protection), borgholzhausen.de (Nov 2025 site restructure broke deep links), liederbach.eu (403 WebFetch-only, curl works — same pattern as pluederhausen.de). Open duplicate-mint/collision list from prior runs (six entries, see 08-29 13:00/15:30 entries) still unchanged — the overdue human consolidation pass remains the single biggest lever for cleaning up the accumulated review queue, and each run's mandatory collision check keeps finding new same-run duplicates on top of the old backlog (now five runs in a row it has caught something), confirming the check earns its keep every time and should never be skipped.