Extend scratch-file workaround to dirs, harden revert-ladder rung 1

[dmccoystephenson]

Summary

• Extends the scratch-file sandbox workaround (Universal rules) to cover scratch directory trees, not just single files: use python3 -c "import shutil; shutil.rmtree(path)" instead of rm -rf (fixes Scratch-file sandbox workaround (create-dev-loop.md:258) doesn't cover directory trees #85 — confirmed on medieval-factions-dev-loop, a session hit the classifier block on rm -rf for a session-created .gradle-local/ scratch dir).
• Hardens the Tests-fix fallback ladder's former rung 1 ("CI-based temporary revert") so a killed session between "push the revert" and "restore the fix" doesn't leave a real broken commit shipping pre-fix behavior on the PR branch, undetected (fixes Tests-fix fallback ladder rung 1 (CI-based temporary revert) has no interruption-safe checkpoint #84):
  • New rung 1: prefer a local, non-pushed revert (git checkout <merge-base> -- <src files> → test → restore) whenever any local anchor can run at all, even partially — it never touches the pushed branch.
  • The CI-based push technique (now rung 2) is reserved for when rung 1 is genuinely unrunnable, and its revert commit message must now be a TEMP: revert <fix-sha> ... MUST be reverted before merge, see <verified-good-sha> marker so a resuming session doesn't need to reverse-engineer the stuck state from a bare red CI run.
  • Added a matching check to Phase 1's open-PR-from-previous-cycle handling: a red final CI run plus a TEMP:-prefixed HEAD commit is now a named signal to reset back to the verified-good SHA and restore, before falling through to generic "CI is red" handling.

Research grounding

No RESEARCH.md finding applies directly — both changes are harness/process robustness fixes (sandbox command classification, and crash-recovery of an in-progress self-review step) rather than an empirical finding about agent behavior.

Test plan

Ran the modified template's Steps 1-4 against a real fixture repo (dmccoystephenson/dans-plugin-manager, cloned fresh, no existing skill — exercised fresh-generation, not update mode) and manually inspected the output:

• Generated skill file has zero unresolved {{placeholders}} (grep -n '{{' <output> — no matches)
• All 13 fenced code blocks survived embedding correctly (26 triple-backtick markers, no leftover escaped backslash-triple-backtick artifacts)
• <!-- template-version: e933cdb --> / <!-- generated-at: ... --> comments present directly below the # <slug>-dev-loop H1
• The new scratch-directory guidance, the renumbered/hardened Tests-fix ladder, and the new orphaned-PR TEMP: check all resolved into the generated output with placeholders substituted correctly (CI, main, repo name, etc.)
• python3 scripts/check_docs.py passes (placeholder/Step-count/link checks)
• Slash-command registration (Step 5) and GitHub repo creation (Step 6) were intentionally skipped for this fixture run — this PR doesn't change those steps, and registering a real skill/repo isn't warranted for a docs-only template validation.

Closes #85
Closes #84

---

This PR description was drafted during a Gardener session (Stephenson-Software/gardener).

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — only create-dev-loop.md modified; every hunk is required by Scratch-file sandbox workaround (create-dev-loop.md:258) doesn't cover directory trees #85 or Tests-fix fallback ladder rung 1 (CI-based temporary revert) has no interruption-safe checkpoint #84 (no unrelated formatting/renames).
• Placeholder-consistency: PASS — no new {{placeholder}} tokens introduced; the diff reuses {{EXTERNAL_SIGNAL_LABEL}} and {{DEFAULT_BRANCH}}, both already rows in the Step 4 table. Confirmed via fixture generation: grep -n '{{' <generated output> had zero matches.
• Escaping: PASS — no new fenced (triple-backtick) code blocks were added, only inline single-backtick code spans, so the leading-backslash escaping rule doesn't apply to this diff. Confirmed the existing 13 fenced blocks in the surrounding template still round-trip correctly through generation (26 triple-backtick markers in the fixture output, zero leftover escaped artifacts).
• Docs: PASS — README's Step count is unaffected (no Step added/removed, only prose within existing Phase 1/3/4 sections); RESEARCH.md needs no update since no finding applies (stated explicitly in the PR body per CLAUDE.md's grounding rule).
• Research-grounding: PASS — PR body states explicitly that no RESEARCH.md finding applies; both changes are harness/process-robustness fixes, not behavioral findings.
• Issue resolution: PASS — Scratch-file sandbox workaround (create-dev-loop.md:258) doesn't cover directory trees #85's suggested shutil.rmtree addition is present verbatim in the Universal rules bullet. Tests-fix fallback ladder rung 1 (CI-based temporary revert) has no interruption-safe checkpoint #84's three suggested actions (prefer local revert first, TEMP:-tagged CI-push as last resort, Phase 1 stuck-mid-revert detection) are all present.
• Sibling structure / sibling renames: no signal this cycle — no new files or renamed identifiers.
• Manual validation: PASS — generated a real skill (dans-plugin-manager-dev-loop.md) from this branch's create-dev-loop.md against a freshly cloned fixture repo (dmccoystephenson/dans-plugin-manager, fresh-generation mode). Zero unresolved placeholders, header comments correctly placed below the H1 with matching SHA, all new prose (scratch-dir guidance, hardened ladder, Phase 1 TEMP: check) resolved correctly with real values substituted.

One thing I found and fixed mid-review (not a FAIL left open, but worth flagging): my first pass at #84 introduced a self-contradiction — a fallback-ladder rung that required "a local anchor that can run" nested inside a ladder explicitly gated on "the local anchor cannot run" (UNVERIFIED), making that rung unreachable as worded. Reworked it as a pre-check before the ladder (a checkout-based retry of the primary stash-and-run experiment) rather than a ladder rung, which also let the CI-based-revert rung keep its original number (1) instead of shifting to 2 — caught and fixed a stale "rung 2" cross-reference in Phase 1 as a result.

Manual validation used the working tree at commit 0fb0046 (this branch's tip). Slash-command registration (Step 5) and GitHub repo/label creation (Step 6) were intentionally skipped for the fixture run — this PR doesn't touch those steps, and creating a real skill/repo isn't warranted for a docs-only template validation.

[dmccoystephenson]

Do-not-auto-merge check: this PR modifies create-dev-loop.md (the template itself), which is on this skill's do-not-auto-merge list — leaving open for human review rather than merging autonomously.

[dmccoystephenson]

Triage note: issues #81 and #79 were also open this cycle but deferred — both require editing CLAUDE.md (agent-loaded config), which needs explicit human authorization per this skill's harness-blocked-operations rule, so neither was attempted. #79 additionally needs a maintainer decision between two resolutions (make gardener public vs. soften the docs) that isn't a docs-sweep call to make.