feat(skills): add pr-swarm parallel PR-loop skill + command

[Mathews-Tom]

Summary

Adds the pr-swarm skill + command pair: parallelizes two or more already-open, file-disjoint pull requests in the same repository to merge-ready at the same time, each driven by its own isolated git worktree and its own independent headless Claude Code session (claude -p) — no shared context/budget, no cross-lane diff or feedback leakage.

Three commits, each independently reviewable:

1. feat(skills): add pr-swarm parallel PR-loop skill — skills/pr-swarm/SKILL.md + references/{launch-mechanics,verification-gates,lane-prompt-template}.md + evals/cases.yaml. Owns the deep mechanics:

   • PR resolution (gh pr view), state/author checks
   • A mandatory changed-file independence check (git diff --name-only on each PR's merge-base) that hard-stops before creating any worktree if two requested PRs touch the same file — never a silent pass-through on the user's own assertion that they're unrelated
   • Branch/worktree short-name inference from headRefName (conventional-prefix stripping, kebab-case, PR-number disambiguation on collision) — never asks the user to name worktrees
   • claude -p launch mechanics: verified flags (--strict-mcp-config, --permission-mode bypassPermissions, --no-session-persistence, --session-id), the lack of a CLI timeout flag (external watchdog required), the zombie-PID liveness-check trap, and an explicitly flagged unverified caveat around background-process survival across tool-call boundaries — documented as an open risk to smoke-test, not asserted as fact
   • Verification gates: stale mergeStateStatus/mergeable (verify with git merge-tree before trusting "conflicting"), the body-only bot-review gap (reviewThreads alone misses findings embedded in a review's raw body with no inline comment), stale statusCheckRollup entries from superseded reruns, and the closed-PR trap (every other field can look healthy on a closed PR)
   • The literal per-lane task prompt template with termination conditions and cross-lane isolation instructions

2. feat(commands): add pr-swarm slash-command wrapper — commands/pr-swarm/COMMAND.md + evals/cases.yaml. Thin argument surface only: /pr-swarm <PR#> [PR# ...] parses PR numbers (with or without #, space- or comma-separated), resolves each against the current repo, and delegates everything else to skills/pr-swarm via a declared command.dependencies entry (mirrors the existing stack-pr/stacked-prs and handoff/handoff same-name command+skill pairing already in this repo). A single PR number is valid input and still runs the full isolated-worktree path — no inline fast path that would silently drop the session-isolation guarantee.

3. chore: regenerate manifest for pr-swarm packages — manifest.yaml via scripts/generate_manifest.py. Adds exactly the two new entries; no other package touched.

Design notes worth flagging for review

• Targets Claude Code's claude -p headless mode, not any other agent CLI — matches this repo's stated Claude Code/Claude.ai scope (see prior issue Docs idea: quick-try link for agent-builder #79 precedent) and the existing agent-builder skill's own headless-mode guidance.
• The independence check is intentionally a hard stop, not a warning — running two lanes concurrently against the same file is the exact failure mode this package exists to prevent.
• references/launch-mechanics.md explicitly documents one mechanic as unverified: whether Claude Code's Bash tool preserves a nohup/disown-backgrounded child across tool-call/turn boundaries. No authoritative source was found either way; the skill instructs checking liveness a minute+ after launch (not immediately) and treats unexpected mid-run death as evidence worth investigating rather than silently retrying. This should be smoke-tested against a real multi-minute run before being treated as settled.

Skill Evaluator Results

skills/pr-swarm (via scripts/evaluate_package.py):

Package: pr-swarm (skill)
  D1 Frontmatter Quality:            20/20
  D2 Trigger Coverage:               18/18
  D3 Structural Completeness:        20/20
  D4 Content Depth:                  18/22
  D5 Consistency:                    12/12
  D6 Compliance:                       8/8
  Overall:                          96/100 (96%)
  Status: PASS

commands/pr-swarm:

Package: pr-swarm (command)
  D1 Frontmatter Quality:            20/20
  D2 Trigger Coverage:               18/18
  D3 Structural Completeness:        16/20
  D4 Content Depth:                  22/22
  D5 Consistency:                    12/12
  D6 Compliance:                       8/8
  Overall:                          96/100 (96%)
  Status: PASS

Zero CRITICAL/HIGH findings on either package.

Checklist

• SKILL.md and COMMAND.md both have valid YAML frontmatter with name and description
• Both package names are kebab-case, under 64 characters (pr-swarm)
• Both descriptions are 200-1024 characters with trigger phrases and a "Use this ... when" clause
• No angle brackets or pushy language in either description
• No secrets, credentials, or internal URLs in any file
• Tested locally: uv run python scripts/validate_evals.py and uv run scripts/generate_manifest.py both pass clean
• All file references (references/*.md) resolve to existing files within skills/pr-swarm/
• No cross-package ../ references; the only inter-package link is command.dependencies: [skills/pr-swarm], matching the existing stack-pr/stacked-prs convention
• Eval cases: skill has 2 positive / 3 negative; command has 4 positive / 2 negative (both exceed the 1+/2+ minimum)
• Skill evaluator score is 70% or above on both packages (96%/96%)
• No CRITICAL or HIGH findings from skill evaluator