de:netzbetreiber-replies

This run is complete. Final report:

Processed all 12 pending reply messages plus, after inadvertently triggering a broader website-hunt batch (a --help probe on hunt_preisblatt.mjs silently ran a real 26-operator crawl instead of erroring — caught and stopped it, but kept the legitimate results), extracted §14a data from those documents too. Net result: 8 full/high-confidence Modul 1/2/3 extractions from mail (Halberstadtwerke, CPM Netz/Evonik, Obernzell, Kempen) plus 7 more from the bonus website hunt (Eirich, Kahl, Waging, Talwerk, Miller), and 7 wrong-document misses flagged with review notes for retry (including a stale 2017 Preisblatt that predates the §14a law, and a municipal hall-rental price list that false-positived on filename keywords). 4 auto-acks, 2 no-document operator pointers now queued for the browser path, and one operator (Schwedt) whose claimed prior attachment never actually reached our inbox.

export_findings.mjs went 487→504 operators tracked, 468→476 complete. Sheet pushed cleanly (504 Findings + 869 Outreach rows). render_reminders.mjs found the mailbox fully triaged with 0 new drafts. The deterministic sender (followups.mjs send --all) was fired detached as required — it completed almost instantly with "nothing selected" (correct, since this pass produced no new followup/reroute drafts). Journal updated with full details, including a script gap in hunt_preisblatt.mjs's staleness filter and the --help footgun for future runs.

fetch_mail.mjs → 12 new messages, all matched. fetch_browser.mjs ran ~9 min again (queue was 45), drained to 5 still-queued + 1 mid-retry, sent 9 new to human queue (rendered pages with no document link, or repeated hop_failed/function-400 errors) — same operators class as the 2026-09-03 05:30 run flagged, budget-for-real-time note still holds.

Accidental bonus haul, worth recording so it isn't mistaken for scope creep next time it happens: ran node hunt_preisblatt.mjs --help to check usage: the script has no --help flag, so the unrecognized flag was silently ignored and it ran a full default batch (--limit 40) against the general operator pool instead of exiting. Caught it via ps aux and killed both the wrapper shell and the node process once ~26 operators had been hunted (not all 40) — real, valid, host-checked website fetches had already happened by then, so I let the output stand rather than discard it. Result: 12 documents retrieved, 14 queued for the browser. Ran pending_docs.mjs immediately after (empty before the accident) and it picked up all 12 as unread documents, exactly as it's designed to for browser-fetched files — so I extracted them as step 2 directs. Lesson for next run: hunt_preisblatt.mjs has no --help; do not probe it that way — read the file or use --dry --mastr <one> instead, or a stray flag silently launches real fetches against up to 40 operators.

Mail-triage extractions (6 full, 1 partial, 2 wrong-document, 4 auto_ack, 2 no-document pointers):

• Halberstadtwerke (SNB977374861035), CPM Netz/Evonik Chemiepark Marl (SNB956923775696), Gemeinde Obernzell (SNB932489723517), Stadtwerke Kempen (SNB943042904373) — clean direct PDF links via fetch_reply_doc.mjs, high-confidence full or near-full Modul 1/2/3 extractions. Halberstadtwerke's document is Modul-1-only (their combined-metering product); the email text said "Modul 3 wird von uns nicht angeboten" so that's recorded as an explicit non-offer rather than a gap.
• GETEC net (SNB995332374861) and Stadtwerke Pritzwalk (SNB922354559020) — operator replied with a link, fetch_reply_doc.mjs followed it and saved a PDF, but it was the general Grundversorgungstarif sheet, not the §14a Preisblatt (both landing pages evidently host more than one document and the script's ranking picked the wrong one). GETEC has no website on file so hunt_preisblatt.mjs couldn't retry; Pritzwalk does, re-hunted, now queued for the browser path. Findings recorded with nulls + review_note rather than guessed from the wrong document.
• Stadtwerke Schwedt (SNB913280322543) — operator's reply claims a Preisblatt was attached to an earlier message on 27.08.2026; that message does not exist in inbox.jsonl (only the 05.08 ticket ack and this one). Nothing to extract from; ran hunt_preisblatt.mjs --mastr on their own site as a stopgap, queued for the browser.
• Stadtwerke Wolfenbüttel (SNB983546347757) — reply says the figures are "auf unserer Internetpräsenz" with no URL. hunt_preisblatt.mjs --mastr found no document by plain GET (likely JS-rendered or behind nav depth 2 didn't reach it), queued for the browser.
• 4 auto_ack (Avacon, Bayernwerk, NordNetz, SH Netz) — ticket confirmations only, action: wait.

Bonus website-hunt extractions (from the accidental batch, 7 full/high-confidence, 5 wrong-document):

• Gebrüder Eirich (SNB953422913031), Gemeindewerke Kahl (SNB962006136537), Gemeindewerke Waging am See (SNB976297675927), Talwerk GmbH/Stadtwerke Kaiserslautern (SNB966823215826) — clean, high-confidence full Modul 1/2/3 extractions straight off their own Preisblätter.
• Gebrüder Miller (SNB926427521488) — full extraction but the PDF's table cells came out of the text layer badly reordered; cross-verified the Modul 2 figure (2,77 ct/kWh) against 40% of the document's own stated base SLP Arbeitspreis (6,92 ct/kWh) before trusting it. Recorded confidence: medium with a review_note flagging the reconstruction.
• 5 wrong-document misses, all from the crawl's link-ranking picking a plausible-sounding file that wasn't the price sheet: Georgensgmünd (a Marktkommunikation contact sheet), Gundelfingen (a Wärmepumpen/Ladepunkt-Anmeldung flowchart), Heikendorf (a Messstellenbetrieb/iMSys price sheet, not §14a), Holzkirchen (a retail Vertriebstarif, "heimat.strom.helden"), Münchweiler (a Bürgerhaus room-rental price list — filename Preisblatt-Buergerhaus-2026-Kunden.pdf scored well on "Preisblatt"+"2026"+"Kunden" despite being completely unrelated; worth a scoring-heuristic look). All six recorded with nulls + review_note; hunt_preisblatt.mjs --retry should get a different file next time per its own dedup logic.
• Gemeindewerke Wildeck (SNB945187645647) — found and downloaded an "endgültiges Preisblatt 2017", i.e. from before §14a EnWG existed (2024). isStale() in hunt_preisblatt.mjs only matches dated filenames with separators (\d{2}[._-]\d{2}[._-]\d{2}); this filename was ab-01012017 with no separators, so it slipped through. Script gap worth fixing: the regex needs a no-separator variant, or these WW00YYYY-style stamps will keep passing as "current."
• Gemeindliche Werke Hengersberg (SNB925878615876) — found a generic §14a-EnWG FAQ (explains "Modul 2 = 60% Reduzierung" as the general rule) with no Hengersberg-specific ct/kWh or EUR figures. Per the normalization rule (a bare percentage without an operator number → null), left it null rather than backing into a number from the generic 60%.

Finish steps: export_findings.mjs → 487→504 operators (+17, all from this run's two extraction batches), 468→476 complete (+8, matching the count of full/high-confidence extractions). push_sheet.mjs pushed 504 Findings + 869 Outreach rows clean. render_reminders.mjs → 0 drafts (mailbox fully triaged, 2 still deferred by OOO, unchanged). followups.mjs send --all fired detached (setsid nohup … & disown); process had already exited by the time checked, log said "nothing selected" — correct no-op, this pass produced 0 new followup/reroute drafts and 0 reminders.

For next run: fetch-queue.jsonl now ~62 entries (grew from the hunt-batch's 14 queued plus this run's own 3 hunts, minus the browser's 1 drained) — expect fetch_browser.mjs to keep taking real wall-clock time. Retry candidates once hunt_preisblatt.mjs --retry is worth spending on: Georgensgmünd, Gundelfingen, Heikendorf, Holzkirchen, Münchweiler, Wildeck, Hengersberg (7 operators, wrong document first try). do_not_contact.txt unchanged (1 entry). No bounces this pass. Standing human-queue list from prior runs (ENGIE Deutschland, Erlanger Stadtwerke JS-render, Heinzelmann/ÜWM successor, Buxtehude GIPS host, Sulzbach empty-body, fetch_reply_doc.mjs Haar ranking bug, Wedel/Metzingen/Energienetze Bayern/Radolfzell/Peißenberg/Netze Calw/ Viernheim-Heddesheim) still open, none touched this pass; add Schwedt (missing referenced attachment) and Wolfenbüttel (no URL given) to it.