feat: introduce AGENT_CMD env var to make runner scripts LLM-agnostic

Summary

Three runner scripts hard-code the claude binary for headless LLM calls. The intent was never to lock all automation to Claude. This issue introduces a single AGENT_CMD environment variable (defaulting to claude) in each script so any compatible headless CLI can be used instead.

Affected Scripts

Script	Hard-coded calls
scripts/autonomous-issue-worker-runner.sh	2× claude --dangerously-skip-permissions --session-id … -p + verify_claude_auth()
scripts/proactive-scheduler-runner.sh	1× claude --dangerously-skip-permissions -p
scripts/dependency-update-sweep-runner.sh	2× claude -p --allowedTools … --output-format text

Approach

Simple env-var substitution. Each script gets:

AGENT_CMD="${AGENT_CMD:-claude}"

All claude binary calls are replaced with "$AGENT_CMD". No flag translation layer — callers must supply a flag-compatible CLI. The --session-id flag (Claude sidebar registration) is guarded behind [[ "$AGENT_CMD" == "claude" ]] since it is Claude-specific.

verify_claude_auth() is renamed verify_agent_auth() and updated to fall through gracefully when the CLI doesn't support auth status.

Full Implementation Plan

See docs/plans/agent-cmd-llm-agnostic-runners.md for the complete detailed plan including exact diffs, comment updates, and SKILL.md prose changes.

Verification

# Smoke test — see what would be called without executing
AGENT_CMD=echo bash scripts/autonomous-issue-worker-runner.sh --help
AGENT_CMD=echo bash scripts/dependency-update-sweep-runner.sh --dry-run

# Full test suite
cd scripts && python3 -m pytest tests/ -x -q

Out of Scope

• scripts/setup-superpowers.sh — claude plugin … is Claude plugin management
• scripts/orion-authorize.sh — claude auth login is Claude auth
• ~/.claude/skills/ path coupling — separate issue
• codex-headless-runner skill — intentionally Claude-specific