feat(skills): add demo-flywheel presenter skill + verification harness

[Manuel-Jentic]

Summary

Adds a presenter tool for demoing the three spec-flywheel flows end-to-end against a local stack, plus a self-verifying harness that proves each flow's real effects landed. It orchestrates the existing skills (import-new-api, contribute-spec-fix) — it doesn't replace them.

Kept un-served (like init-design): it's a presenter tool, not an agent capability, so there's no SERVED_SKILLS entry, no cli//src/ mirrors, and no drift-test obligation (drift test stays green).

The demo (one API's lifecycle, in 2 → 1 → 3 order)

flowchart LR
  F2["Flow 2\nimport a NEW API"] --> F1["Flow 1\nimprove it via an overlay\n(confirm makes it live)"] --> F3["Flow 3\nreact to an upstream change\n+ rollback (teardown)"]

Loading

Flow 1 needs an already-catalogued API, so running Flow 2 first (which creates one) makes the whole demo a single coherent subject: import → improve → live → react/rollback.

What's here

• skills/demo-flywheel/SKILL.md — the narrated walkthrough with an explicit 4-gate human-in-the-loop protocol (agent approval, overlay :confirm, any public gh action, rollback), auto-proceeding on everything local/reversible.
• scripts/flywheel_verify.py — an httpx PASS/FAIL harness (--flow 2|1|3|all). Bootstraps its own org-admin, promotes imported drafts, polls the overlay for confirmed_revision_id, and exits non-zero on any failure (so it doubles as a smoke test). Flow 3 self-skips if the fixture isn't up.
• config/local-sqlite-demo.yaml — a non-destructive full-copy config that adds only what Flow 3 needs: the fixture manifest_url, update_check_interval_seconds: 1 (never 0, which is the kill switch), and 127.0.0.0/8 egress so the loopback fixture is probeable.

Safety / blast-radius

A real `[AUTO]` issue on `jentic/jentic-public-apis` triggers a live LLM-backed workflow that spends the repo's Bedrock secret and opens a non-draftable PR to `main`. So the demo defaults to a `--no-github` dry-run or a presenter-owned fork, makes the Flow-1 human PR a draft with a `[DEMO — do not merge]` title, uses a fresh `RUN_ID` per run (to dodge the existence-gate STOP and the duplicate-version hard-fail), and ends with a mandatory teardown (rollback + close/delete PR/issue/branch). The harness itself never touches GitHub.

Test plan

Verified end-to-end against `make start-app-sqlite` (with the demo config) + the Flow-3 fixture:

• `python scripts/flywheel_verify.py --flow all` → 21/21 PASS (import + promote + listable/live; overlay submit → confirm → served-spec fix → rollback reverts → deprecated; fixture bump → refresh → `update_available` flips → re-import clears).
• Flow-3 skip path exits 0 when the fixture is down.
• `ruff` + `ruff format` + `mypy` clean on the harness.
• `tools/skills_sync.py --check` and `tests/arch/test_skill_drift.py` green (un-served skill imposes no obligation).

Reviewer note: please use squash + merge.

Made with Cursor