chore(ci): self-heal corrupt runner checkout in scrub-refs step

[briansrls]

Problem

Shared self-hosted runners (srv1/srv2) persist the repo checkout between jobs. When a runner's checkout gets a ref whose object is missing, every subsequent Checkout on that runner dies — observed on srv2-23:

fatal: bad object refs/remotes/origin/session/fierce-ram-162

This reds CI for any PR that lands on the wedged runner (recently hit #4918's layering_imports and others). The existing "Scrub dangling remote refs" step (which update-ref -ds dangling refs) does not clear it when update-ref -d can't remove the ref (e.g. corrupt packed-refs) or when git for-each-ref itself errors.

Fix

Harden the scrub step in all 13 jobs: after the dangling-ref pass, if the ref store is unreadable or any ref tip object is still missing, rm -rf the repo dir so actions/checkout fresh-clones. This self-heals the wedged runner on its next job, before checkout — taking the operator off the checkout-repair axis (no manual runner re-provision needed for this failure mode).

Safety — the rm -rf is precise by construction

It fires only when git for-each-ref errors or a ref tip resolves to a missing object. A healthy or empty checkout is never removed.

Verified by execution (local repro):

• ✅ healthy repo (incl. refs/remotes/origin/session/* refs) → kept
• ✅ fresh repo with no refs → kept
• ✅ repo with a ref to a missing object (the srv2-23 case) → detected & removed
• ✅ normal case where update-ref -d clears the bad ref → no rm -rf (last-resort only)

Note

This PR will itself sit red on v4_claim_witness_corpus_b until the separate placement_supply.dag Nat-multiply main-red is fixed (snappy's lane) — that's the corpus_b axis, orthogonal to this checkout-axis fix. It can merge once main greens, or be merged as an infra fix.

🤖 Generated with Claude Code

[briansrls]

The single failing check — v4_claim_witness_corpus_b — is an inherited fleet-wide main-red, not caused by this PR. It fails on:

dsl/std/placement_supply.dag:41:59: error: function '<expr>' not found in scope
error: P-PROBE-CF-IMPORT repro failed but without expected Option-not-found diagnostic

This is the placement_supply.dag Nat-multiply resolve break (exposed by #4883's probe_selector_compute_fabric_import_repro witness) that currently reds shard B on main itself (ci.yml run on f109f463d0 is red on the same job), so every PR inherits it. It is being fixed in the compute_fabric/measure lane (snappy's Nat-mul fold); orthogonal to this PR's checkout-axis change.

This PR is ci.yml-only and is otherwise healthy: layering_imports, fmt, doc_refs, infra_isolation, timeout_budgets, v4_claim_witness_corpus_a, and all v4_lens_* jobs pass — confirming the workflow parses and the hardened scrub + checkout run cleanly. No fix is applicable in this PR; it clears on main-green + rerun (and its own scrub-hardening keeps the rerun off any wedged-checkout runner).

— sent from swift-stag-552