Fix autoloop scheduler starvation: clone repo-memory before pre-step, deterministic tiebreak

[Copilot]

The autoloop pre-step reads program state from /tmp/gh-aw/repo-memory/autoloop/, but gh-aw's built-in repo-memory clone runs after the pre-step (and into a different directory). Every program looks like a "first run", and the implicit insertion-order tiebreaker permanently starved build-tsb-pandas-typescript-migration behind perf-comparison for 8+ days.

Changes to .github/workflows/autoloop.md

• Clone memory/autoloop before the scheduler runs. New first step in steps: clones the branch into /tmp/gh-aw/repo-memory/autoloop/ so the existing Python pre-step actually finds state files. Idempotent (skips if .git already present); falls back to an empty dir when the branch doesn't exist (genuine first run).

• Auth via http.extraHeader instead of token-in-URL to keep the token out of process listings and any logged remote URLs.

• Deterministic tiebreaker for the due-list sort. Replaces due.sort(key=lambda p: p["last_run"] or ""), which fell back to insertion order when last_run was absent. New key:

  • never-run programs sort before ever-run programs;
  • among never-run, shortest schedule wins, then alphabetical by name (programs without a parseable schedule fall back to +inf);
  • ever-run programs continue to sort by oldest last_run, then name.

  Each due entry now carries schedule_seconds to support this.

def _due_sort_key(p):
    if p["last_run"]:
        return (1, p["last_run"], p["name"])
    sched = p.get("schedule_seconds")
    return (0, sched if sched is not None else float("inf"), p["name"])
due.sort(key=_due_sort_key)

Effect

After merge, the pre-step should log build-tsb-pandas-typescript-migration: last_run=2026-04-12T11:15:07Z, iteration_count=230, that program wins the next selection on most-overdue, and ordinary last_run ordering takes over. The tiebreaker change is defense-in-depth: even if state goes missing again, every 30m beats every 6h instead of insertion order deciding.