Refresh All Prisms

Force a recompute of every live prism, bypassing the freshness gate.

The gate only opens once every granted table's table_max_date advances, so a compute-side change can't reach the public page while sources are idle — the last-good graphs keep serving indefinitely. This is the operator escape hatch for that case.

Queues one task_refresh_prism per prism rather than computing inline: each is a real purchase against the warehouse, and 20+ of them would run well past any HTTP timeout. Outcomes land in the recompute-attempts log like any other compute. Dispatched by name because prism_tasks imports from this module.