feat(multi-issue): add Phase 0 backlog triage + edit-only-agents note (v1.2.0)

[wan-huiyan]

What & why

Field-tested overnight-multi-issue-implementation against a 36-issue, partly-stale, no-CI backlog (user: "tackle them all, some might be outdated"). The skill's preconditions assume pre-validated issues — that broke. This adds the missing front-end + one orchestration note. (The repo had already absorbed the red-baseline "identical failing set" gate and the shared-file sequential-resync pattern — those are not re-added.)

Changes

1. ## Phase 0 (conditional): triage a large or possibly-stale backlog (new section) + a "When to use" pointer. The load-bearing insight is the dismissal asymmetry: a wrong still_valid is caught at implementation, but a wrong already_fixed/outdated silently drops real work. So the verdict schema is biased against dismissal (dismissal requires PR# + file:line), and an adversarial pass verifies only the dismissals. Ends on a user-ratified cut-line — never auto-close. Self-contained; references the standalone skill stale-backlog-triage-adversarial-verify.
   • Real result that motivated it: 36 issues → adversarial pass flipped 1 wrongly-dismissed (file deleted but 3 capabilities silently un-migrated) + confirmed 1 done; only 1/36 was actually done. Naive implementation re-does merged work; naive dismissal drops a live issue.
2. Parallel implementers — keep git in the controller (Phase A/B note). For concurrent file-disjoint PRs: agents edit + test only in isolated worktrees; controller owns all git. Sidesteps subagent-bash-cd-wrong-worktree / ...-branch-ref-froze-stranded-commits / ...-reports-complete-but-PR-unmerged.
3. Version → 1.2.0 across SKILL.md frontmatter + plugin.json + marketplace.json (also fixes the pre-existing 1.1.0-SKILL.md vs 1.0.0-manifest drift).

Validation

python3 .github/scripts/validate_plugins.py → OK: marketplace + plugins valid (0 warnings). JSON lint clean. Versions consistent.

🤖 Generated with Claude Code