feat(protocol): add universal agent session entrypoint

[localai-bot]

Row

SESSION-ONBOARDING-UX — governance/session-entrypoint row.

Before starting

• Issue/PR search and existing claim: no matching issue or PR; the helper role marker claims SESSION-ONBOARDING-UX in this worktree.
• Roadmap or matrix row, plus scripts/ready-for-helper.py result when applicable: governance-only protocol work specified by .agents/specs/session-onboarding.md; python3 scripts/ready-for-helper.py --check-local returned OK: LOCAL_ONLY readiness structure validated; live claims were not verified and no helper queue was advertised.
• Exact current-code and test/evidence anchors inspected: scripts/agent-role.py, scripts/agent-onboard.py, scripts/agent-preflight.sh, .agents/workflow.md, .agents/prompts/operator.md, and their mutation suites under tests/scripts/.

What changed

Adds a universal, non-interactive scripts/agent-start.py session entrypoint. It emits a source-owned ASCII welcome only when no role and no explicit intent exist, explains Helper/Operator/Looking roles, reports truthful inherited state, and routes the agent through declaration, rerun, preflight, and resume. It also strengthens POL-REVIEW-NO-REPAIR: every actionable in-scope review failure must return to a fresh implementer, run focused and full gates, and receive a fresh scoped review until PASS; retry budgets cannot terminate correctable findings.

Evidence

• scripts/agent-preflight.sh && git push -u origin row/SESSION-ONBOARDING-UX-IMPL exited 0 with All gates green; the push executed only after the gate passed.
• python3 -m unittest tests.scripts.test_agent_start tests.scripts.test_agent_onboard tests.scripts.test_policy_contract tests.scripts.test_check_prompt_contract tests.scripts.test_check_protocol_consistency ran 156 tests: OK.
• A fresh reviewer inspected immutable 3337bda3, ran eight targeted scratch mutations plus one full preflight, and returned PASS with no findings.
• Same-change documentation obligations are recorded in docs/STATUS.md and docs/BENCHMARKS.md; docs/FEATURES.md is unchanged because no feature/model/backend/quantization surface moved.

Speed claims

• This PR makes NO speed claim (helpers may not: the GPU is operator-serialized and a contended benchmark is void), or
• the operator ran the numbers under ${GPU_LOCK} and they are recorded in docs/BENCHMARKS.md with the repro recipe.

Honest gaps

• No inference, GPU, or performance behavior changes.
• The entrypoint deliberately prints actions and never claims roles, creates worktrees/PRs, acquires locks, or mutates gates itself.
• Untracked SDD execution metadata (.sdd-state.json, docs/spec/) is intentionally excluded from the branch.