Fan-out 2026-09-12: handoffs, Telegram, ICM, scanbot review

What changed

• Handoff stalls fixed at the root. The runner copy shipped to Nano was arriving truncated because the source file was rewritten while three handoffs staged, and cat > accepts a short stream and exits 0. The ship now goes inside a tar stream, then the receiver's parse result and md5 are compared to the local copy and the handoff dies on any mismatch. Proven against real Nano (parse OK, md5 match) and with a 4-case stubbed test (clean pass, cut pipe, truncation, byte tweak).
• Every finished task now ships a phone-readable summary. New generator summary-html.sh plus template, wired into the receiver runner tail and the handoff card. Body is simple markdown, output is one HTML under 60 KB, and tg-publish renders it to a PDF that opens inline. This page was produced with it.
• Telegram noise closed at the senders. 13 M2 sender scripts now route through the classed helper and honor the mute conf: progress and watcher chatter muted, done and alert pass, zero parse_mode, no .md path or bare filename payloads. Three older copies on Nano upgraded the same way. One ungated path remains: the screen-preview photo send.
• ICM registry clean. Lint errors 5 to 0. Five real skill ids added so dangling references resolve. New block B108 (unverified stdin ship) recorded. A block written on Mini today by another lane merged fleet-wide. Seven shared block numbers stay as-is with a dry-run renumber proposal.
• Computer-use layer hardened. Four verified defects fixed: preflight accepts the long profile names, the Accessibility probe retries instead of flapping, "apple mail" now routes computer-use first, and the BLOCKS layer prints NOT AVAILABLE instead of nothing. Four reusable drive scripts extracted: chrome-js, safari-js, real-click, storyline-advance.
• Nano ICM card repaired. KNOWLEDGE and COMMITMENTS were dead on Nano (old vault-search shim without JSON output, no registry directory). Full v2 tree pushed, commitments rebuilt, all four layers answer.
• Scanbot L3/L4 adversarial review: DO NOT PROMOTE. 4 critical, 8 major, 10 minor. Worst: a redacted derivative of one document can be bound to another document's approved artifact. The promotion target is also wrong: the live caller is the taxbot scripts dir, not ops/scripts/scanbot. The fix lane stopped itself because another session is rewriting that build dir right now.

Evidence

Next step

• Scanbot: decide who owns the build dir. The other session has C1 and a different M4 done; C2, C3, C4, M1, M5 are still open and C2/C3 need exclusive ownership of db.py. Re-point the cutover at the live taxbot scripts dir.
• Nano vault index is 20 days old. Rebuild with the vault indexer on Nano when convenient.
• Screen-preview photo send needs a photo mode in send-tg or a switch to document delivery to be mute-governed.

Open decisions

• GO or amend the seven-pair block renumber proposal (dry-run script refuses without --apply).
• Inbound Telegram poller owner: Nano tg-poll or Mini runtime, exactly one.
• Mini's watchkeeper and Nano's session-close script diverged from M2 and were left untouched.
• Earlier items still yours: council plist bootstrap, stray-dir delete GO, cutroom installs GO, Mini tailnet re-auth.

Packet fields