ClaimNext charges an attempt to every claim — right for a delivery that was TRIED, wrong for one that was HELD. park/parkCausal settled through Release, which never gave the increment back, so a kill switch held for ~40s exhausted an 8-attempt budget and the first retryable failure after the switch cleared poisoned instead of retrying (FOLLOWUPS "DEFECT, NOT DONE", 2026-08-14). ReleaseParked is Release's fenced sibling built from one shared template (the fence is spelled once); its single difference is attempts = GREATEST(attempts - 1, 0). The claim-token fence is what makes the give-back safe: only the claim that added an increment can subtract one, a stale park matches zero rows, and a terminal row is untouched. Real failures (releaseOrPoison) and held settlements (settleLater) keep Release — those increments are earned. The causal wait's outcome remains wall-clock-decided; a regression guard pins both facts together. Tests: outer acceptance (switch held past the cap, then one 503 → retry, not poison), store fencing/net-zero suite, all three park sites attempt-neutral. DEPLOY.md §5, README, and FOLLOWUPS rewritten to the fixed behavior; the residual write-load cost note stands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>