safe_outputs: always use the same checkout layout as the agent job (not only for target-repo: "*")

[dsyme]

Summary

The safe_outputs job should always use the same checkout layout as the agent job. Today it only mirrors the agent layout when safe-outputs.create-pull-request.target-repo (or push-to-pull-request-branch.target-repo) is the wildcard "*". For any other configuration it falls back to a divergent, single-repo layout. This divergence is unnecessary complexity and is a source of bugs.

Current behavior

buildSharedPRCheckoutSteps in pkg/workflow/compiler_safe_outputs_steps.go branches on targetRepoSlug:

• target-repo: "*" → buildMultiRepoCheckoutSteps, which mirrors the agent job: default repo checked out to the workspace root, plus every checkout: entry checked out to its configured path:, then matching Fetch additional refs steps.
• specific slug (e.g. github/github) → single-repo path: checks out only that one target repo to the workspace root (no path:), with the base branch as ref:, followed by Configure Git credentials and a single fetch step.
• no slug (same-repo) → checks out the workspace repo to root.

So the layout an author gets depends on the target-repo value, and only the wildcard case matches what the agent job actually checked out.

Why this is a problem

1. Layout drift between agent and safe_outputs. The agent operates against one on-disk layout (e.g. workflow repo at root, github/github at ./github); the safe_outputs job applying the patch/creating the PR sees a different layout (e.g. github/github at root). Handlers then have to special-case which directory the repo lives in.

2. The two-cross-repo case is unsupported. Consider a workflow with two checkout: entries, e.g. org/a → ./a and org/b → ./b, and create-pull-request.target-repo: "org/a" (a specific slug, not *). The agent job checks out both repos at their paths. The safe_outputs single-repo path checks out only org/a, to the root, ignoring org/b and ignoring the path:. Any handler that expects the agent's layout (e.g. to read context from ./b, or to find org/a at ./a) breaks. There is no reason the specific-slug case should behave differently from the wildcard case here.

3. It caused a real failure. With a specific target-repo, the single-repo path checks out the target to root and then runs Configure Git credentials (which rewrites origin to an embedded-credential URL) before the fetch step, and omits the "Clear partial clone markers" reset that the agent path emits. The agent job's identical fetch — run against a clean origin in a subdirectory — succeeds, while the safe_outputs fetch fails with fatal: couldn't find remote ref refs/heads/master. Unifying on the agent layout removes this class of bug.

Proposed change

Always generate the safe_outputs PR checkout steps using the same layout as the agent job, for all target-repo values (specific slug, *, or unset) — i.e. always take the multi-repo / agent-mirroring path. The handlers already locate the correct repo via findRepoCheckout() / the checkout manifest, so the wildcard machinery generalizes to the specific-slug case.

Considerations to preserve:

• safe_outputs must check out the base branch (not github.sha) for PR creation — keep the extract-base-branch step and base-branch ref: for the root checkout.
• Keep the run conditions (create_pull_request / push_to_pull_request_branch).
• Keep git-credential configuration for all checked-out repos so pushes work.
• Mirror the agent's partial-clone-marker reset and fetch-ref steps so cross-repo fetches behave identically.

Acceptance criteria

• safe_outputs PR checkout layout is byte-for-byte equivalent (modulo base-branch ref: and run conditions) to the agent job's checkout layout for: same-repo, single specific cross-repo, two-or-more cross-repo, and wildcard * configurations.
• The two-cross-repo / specific-slug case checks out every checkout: entry to its path:.
• No regression in PR creation / push-to-branch for same-repo and cross-repo workflows.
• The couldn't find remote ref fetch failure no longer reproduces.