CORRECTION: 2026-07-18/19 GH-auth-blocker DM spam continued ~22h past prior fix-and-close (995153f2 was premature)

Task 995153f2 closed at 2026-07-19T10:24:50Z with result "Root cause: auto-surfaced incomplete turn errors. HEARTBEAT.md update is sufficient." That conclusion was wrong. Live Slack history review during the 2026-07-20 04:10 nightly reflection shows the SAME false GitHub-auth-blocker claim (alternating broken/fixed/broken) continued at roughly the same cadence for ~22 more hours after the closure and after the 05:45:47Z apology DM, with spam-pattern messages at approximately 2026-07-19T11:55, 14:39, 17:54, 18:24, 19:10, 20:09, 21:09, 22:24, 22:54, and 2026-07-20T00:24, 01:09, 02:55 UTC -- roughly 12 MORE false-blocker sends after the fix was declared sufficient, plus ~5 more false-resolution ("actually it is fixed now") messages that also proved wrong within a few hours each. Total scope across the full arc (2026-07-18T16:39Z to 2026-07-20T02:55Z) is ~21 false-claim messages, not the 12 originally documented -- roughly double, and continuing literally up to ~1h before this correction. Live verification during this cron (eval gh-fleet-token.sh + gh auth status + gh pr view) again shows zero issues, consistent with every prior spot-check -- the claim is never true when actually tested, yet keeps re-firing. This strongly suggests the send path is NOT going through normal heartbeat reasoning that would read HEARTBEAT.md (which now contains explicit, incident-named language telling the agent not to do exactly this) -- it is more likely an auto-surfaced error-recovery or fallback message path in the agent runtime that bypasses HEARTBEAT.md/CHECKLIST-DM.md entirely. This is now an infrastructure-level hypothesis, not an agent-reasoning-discipline hypothesis, and needs investigation with full gateway/session log access outside a reflection-cron sandbox, likely requiring Victor (fleet infra owner) rather than further prose edits to agent-side markdown files. Recommend: (1) do not mark this class of task complete again without a >=24h silent-window verification; (2) escalate to Victor for gateway/runtime-level investigation of auto-surfaced error messages bypassing send-gates; (3) if a 4th recurrence happens, treat it as confirmed infra bug requiring a runtime fix, not another markdown edit.