Recover approval-gate backlogs - #999

Summary

Fixes the systemic approval-gate black-hole described in #998 across the daemon registry, API, and broker.

• re-prompts aged pending approvals at 24h and 72h, with durable bounded counters and no new inbound message required
• exposes fleet-wide and per-agent pending/undelivered backlog diagnostics
• settles approval-request state on every registry approval transition and continuously flushes approved-but-stranded rows, including startup/migration leftovers
• sends an immediate loud alert when a pending chat holds a message from an already-approved principal of the same agent

Implementation

• adds durable aging_reprompt_count and high_signal_alerted_at migration columns
• mirrors the incident SQL oracle by grouping undelivered pending_messages on (agent_name, chat_id) and joining both gate status and sender approval status
• adds GET /system/approval-backlog, agent health checks.approval_backlog, and backlog detail on the pending-messages response
• runs approved-backlog reconciliation at startup before pollers and continuously in the approval maintenance loop
• serializes each approval flush and checkpoints a row only after the streaming handoff succeeds, so offline startup reconciliation cannot silently mark a dropped message delivered
• preserves the existing transport retry and new-message re-notification behavior

Validation

• focused approval/registry/broker/grandfather suite: 207 passed
• repo-wide Ruff: passed
• frontend i18n extraction, tests, and production build: passed
• full Python suite: 4,607 passed / 3 skipped; two unrelated failures inherited live-host PINKY_DREAM_TRANSPORT, PINKY_AUTH_DENY_DEFAULT, and PINKY_SHARED_MCP settings
• both environment-contaminated failures pass when rerun with those host overrides removed (2 passed)

No release or deploy is included; this rides the next natural cut.

Closes #998