de:wasserversorger-discover

All files are valid and the journal entry is written. Here's the run report:

Resumed a batch that a prior invocation had left mid-flight (9 of 60 units done, 60 assigned). Two subagents finished the remaining 53 units — 43 fully resolved, 10 partially resolved, 0 fully abandoned — adding 74 new companies and 691 municipality links, then I merged 3 real cross-shard duplicate companies (kastorf, güstrow, ludwigslust water Verbände) found by the end-of-run domain-collision check. Coverage rose from 3,324 to 3,898 municipalities claimed (34.75% of the German population, +0.85pp), the phase-1 queue (fragmented units still worth block-claiming) shrank from 367 to 315 units, and map_roster.mjs merged cleanly with 0 ambiguous/unmatched rows. Journal updated with the full details, including a newly-caught id-range-collision failure mode and two new unreachable-host findings (docplayer.org, web.archive.org) for the next run.

Same shape as the 18:00 run above: on arrival, shards/batch-2026-08-20-part-j.json and -part-k.json (60 units, generated ~00:00) plus partial roster.part-j/k.jsonl and roster-links.part-j/k.jsonl already existed with no journal entry — a prior invocation had picked the batch and gotten 9 of the 60 units fully done (plus a few more partially started) before stopping unmerged. Used the same AGS-overlap-against-existing-links check to find the exact remaining work per unit (this is now the third run in a row this has worked reliably) and dispatched 2 subagents — one per pre-existing shard/id-range — to finish it, this time running them in the foreground (run_in_background: false, both in one message) rather than background, specifically to honor resumability rule 4 (a background subagent's work dies the moment the parent's turn ends, and two prior runs' journal entries recorded exactly that failure). Both instructions explicitly forbade the subagent from spawning further sub-agents or writing outside the campaign files, quoting the 18:00 run's /tmp incident by name — neither subagent repeated it this time.

Shard J (26 remaining units): 16 fully resolved, 10 partially resolved, 0 fully unresolved. 356 link rows / 40 company rows appended. Confirmed two more instances of the running "WebSearch AI-summary hallucination" failure mode, both caught before writing: a summary invented a Zweckverband-Altshausen link that its own Satzung contradicted, and a summary misplaced one Neverin village that the primary Verband's own member table placed correctly — in both cases the primary source overrode the synthesis. New unreachable host: docplayer.org (ENOTFOUND — not a block, just gone; use the search-snippet text instead of trying to fetch it).

Shard K (27 remaining units): all 27 finished — 25 genuinely researched, 2 (Hagenow-Land's last member, all of Trave-Land... on closer look Trave-Land was not actually pre-resolved, see below) found already covered. 335 link rows / 34 company rows appended. New finding: web.archive.org is now hard-unreachable to WebFetch in this environment (not a site-side block) — the wayback-machine workaround the 03:00-ish predecessor run used for wbv-sude-schaale.de is no longer available; that domain now needs the live site or a different mirror. No pdftotext/python3 in this environment (confirmed again) — one scanned Satzung (Landstuhl) had to fall back to WebSearch synthesis at medium confidence instead of a direct read.

Company-id range collision, caught mid-run by shard J itself: both subagents were told disjoint ranges (J: 710+, K: 810+) from a snapshot taken before dispatch, but shard K had already raced ahead past 810 into the 800s by the time J started minting — J found real overlap (including one true double-use of a single id for two different companies) and self-corrected by renumbering its own new companies into 900–931 before more damage was done. Lesson for next run: don't hand out a fixed id-range per shard in the dispatch prompt — tell each shard to read the current max id across every roster.part-*.jsonl immediately before its first mint, and check again periodically if the run is long, since a sibling shard's range is a moving target while both run concurrently.

Domain-collision check across all shards after merge found 4 collisions, 3 real: wbv-kastorf.de (wv-de-362 vs wv-de-703 — same Wasserbeschaffungsverband Kastorf, independently found by an earlier run and by shard J; 703 had the better primary-source evidence — own Versorgungsgebiet page vs a third-party summary — so it was kept, 362's 2 links repointed and its row dropped), waz-guestrow.de (wv-de-634 vs wv-de-801, both from this same Land/Kreis being worked by two different shards independently — 634 had the more directly-sourced record so it won, 801's Sternberger-Seenlandschaft-specific finding was folded into 634's notes before dropping 801 and repointing its 21 links), zkwal.de (wv-de-670 vs wv-de-802, same shape — 670 kept, 802's Ludwigslust-Land- specific carve-out detail folded in, 7 links repointed). The 4th, emkendorf.de, is the already-confirmed legitimate 3-way share from the 18:00 run — reconfirmed, not a bug. All three real merges followed the established pattern: keep the better-evidenced id, fold any genuinely new information from the loser's notes into the winner, repoint links, drop the duplicate row — never silently pick one and discard the other's findings.

Net for the run: 51 of 53 remaining units in the batch actively researched (2 came pre-resolved from concurrent work). Roster: 254 unique companies across roster.part-*.jsonl (32 pending_domain). map_roster.mjs: 3,667 total link rows, 0 ambiguous, 0 unmatched (3,484 pre-seeded/exact, 165 exact_norm, 15 land_scoped, 3 expanded_qualifier). coverage.mjs: 3,324 → 3,898 municipalities (+574), population 33.9% → 34.75% (+0.85pp). build_gv_layer.mjs: phase-1 pool (members_residual ≥ 5) 367 → 315 units — well above zero, so phase-1 continues; do not stop. New queue top: Mittelholstein (16 residual, down from 26 but still the largest single gap — mostly small century-old single-village Wassergenossenschaften that resist verification within a normal budget; worth a dedicated pass or accepting a lower confidence bar for these), Elm-Asse (12, fresh), Trave-Land (12 residual — still not fully cracked despite two runs' attention now), Seenplatte (12, fresh), Schwarzatal (10, fresh, untouched by any run so far), Parchimer Umland / Leezen / Franzburg-Richtenberg / Neustrelitz-Land / Boizenburg-Land (10 each, all fresh MV/SH units).

No new CAPTCHA/anti-bot blocks; no form submissions or contact attempts made (research only, per country rules).