Sync with pstack 0.15.x (0.15.6 -> 0.15.21), pi adaptations preserved

[qapquiz]

Syncs qstack with everything upstream pstack shipped since the Sep 27 port: cursor/plugins pstack 78f46da..997ffc7 (0.15.6 → 0.15.21, ~25 commits, 88 files).

All pi adaptations are preserved and re-applied on top of upstream's changes: pi -p spawns instead of the Task tool, the pstack models section of ~/.pi/agent/AGENTS.md instead of Cursor rules, /skill: invocation names, agents as --append-system-prompt briefing files, and no hardcoded model defaults.

New skills (pi-adapted)

• /skill:correct — mine repeated agent mistakes, fix each at the highest level that works
• /skill:ketchup — catch-up on everything since your last message, with executive summaries per action item
• /skill:benchmark-checklist — vet a perf number before reporting or acting on it
• /skill:poteto-help (+ prompting/recipes references) — routed through qstack's install and config surfaces
• principle-explain-the-number — rides along via the leaf-naming convention

Behavior changes ported

• poteto-mode: the chat that receives /skill:poteto-mode runs the playbook; poteto-agent is a delegate only (upstream 2bd5dd4). Fresh spawns by default; resume only when a spawn's state is costly to move. New benchmark-checklist trigger. Bundled-skills path note for poteto-agent prompts (364a77e).
• setup-pstack: unlimited now sets max thinking instead of leaving entries as-is; large matches the new defaults (df58112).
• Autopilot playbooks: hourly audit ticks instead of ~30 min (tmux/shell watcher kept as the pi mechanism), owners push after every verifiable unit, merge-prep gains merge-tree + CI-config-path checks, a fresh owner takes each next queue item, the /goal mechanism is dropped (as upstream did), multi-phase tick prompt simplified.
• arena: candidates must name the alternatives they rejected in their rationale.
• create-verification-skill: never weaken a journey to make it pass; readiness means the app is usable (c49b37a).
• Scripts: watch-pr paginates review threads (#484), plan checks respect fenced code blocks (#536), worktree-audit checks MERGED against the default branch (#574). Plus upstream's new check-plan.test.ts.

Docs/README: investigation-first understand guide, 24 principles, Cursor-specific additions (Custom Modes, cloud subagents, Projects) re-expressed for pi (sticky mode, own worktree, coordinator chats), and /skill: conventions applied throughout.

Verification: bun test in skills/poteto-mode/scripts — 63 pass, 0 fail on this branch. On main the same suite has 4 failures that these upstream script fixes close.

Deliberately not ported: upstream's plugin.json Agent Plugins v1 manifest (qstack's pi manifest stays in package.json), .cursor-plugin/plugin.json, and the benny automations pack (unchanged upstream; still Cursor-only).

Upstream refs: #484, #486, #494, #502, #511, #536, #538, #539, #561, #570-#584.