fix(quota): deduplicate vision audit output

[huangruiteng]

Summary

• replace four repeated full vision continuation audits in default quota should-run output with bounded decision anchors
• preserve required, decision, recommended_action, closure flags, and deterministic JSON refs
• expose the original repeated payload through --include-vision-audit-detail

Real-scale result

• PR fix(quota): bound CLI todo diagnostics #2510 default: 65,702 bytes / 1,384 lines
• this PR default: 51,256 bytes / 1,169 lines
• reduction: 14,446 bytes (22.0%)
• combined full-detail mode remains available at 102,878 bytes
• selected_todo, scheduler_hint, and the required/decision/recommended-action anchors are exactly equal across default and detail modes

Qualification

• focused pytest: 11 passed
• pinned Ruff: passed
• risk-based premerge: 17/17 checks passed, no failures or manual holds
• public/private boundary scan: clean

Dependency

This is intentionally stacked on #2510 so the preceding todo-summary compaction stays reviewable. Retarget to main after #2510 merges; do not merge this PR first.

[huangruiteng]

压缩后会丢失 required_before_closeout、vision_gap_judge、agent_judge_instruction、evidence/registry read instruction 等字段。测试错误地把 CLI channel 伪造成完整 audit，因此没有捕获。