docs+skill: onboarding UX (Node 22 prereq, version pin, explicit model selection)

[tangym]

Summary

Three small onboarding-experience fixes for wire-assert-ci. All three cover places where the current flow silently fails or silently picks a suboptimal default.

Changes

Area	What changed	Why
README.md — Option 1 install	Add Requires Node.js ≥ 22 note and cross-link to Option 2 as fallback	The skills CLI (skills@1.5.22) imports styleText from node:util, added in Node 22. On fresh devcontainers running Node 18/20 (Debian bookworm default) users hit SyntaxError: The requested module 'node:util' does not provide an export named 'styleText', which reads as unrelated to Node version.
wire-assert-ci skill — new ## assert-ai version pin section (all 3 mirrors)	At workflow-authoring time, resolve latest stable from PyPI via curl -fsSL https://pypi.org/pypi/assert-ai/json | python -c ... and pin the exact version into assert-ai-version. Users own subsequent bumps.	The action's gate uses paired-binary McNemar across baseline and PR runs. Pinning is required so drift in assert-ai (prompt formatting, judge behavior, tokenization) doesn't get attributed to code drift and fire false regressions.
wire-assert-ci skill — new ### Model selection section (all 3 mirrors)	Skill must ask the user for judge / tester / default roles in one prompt before writing any behavior YAML. Use LiteLLM-style names (azure/<deployment>, openai/gpt-5.4, anthropic/claude-4-sonnet). Verify provider-env secrets match. Write every role explicitly. Never let pipeline.judge.model silently fall back to default_model. Never write tester: null without explicit single-turn confirmation.	Available models are user-specific (Azure deployment names, provider access). Guessing a default silently downgrades the paid-provider gate to whatever small model happens to be the fallback.

Skill edits are applied identically across the three mirror files: SKILL.md (Claude), wire-assert-ci.prompt.md (Copilot), assert-ci.mdc (Cursor).

Not included

Earlier iterations of the skill also carried a YAML-emission mitigation (instructing the skill to hand-format multi-line strings with | block scalars). That's dropped here — the correct fix lives upstream in assert-ai init itself: responsibleai/ASSERT#314. Once that lands and the pinned version rolls forward, behavior YAMLs will be emitted with block scalars and verbatim unicode automatically.

Testing

• Ran the updated Option 1 instructions on a fresh Node-18 devcontainer: hit the styleText error as expected, then Node-22 install path worked.
• Walked the skill end-to-end via Copilot chat against the demo repo: asks for model choices before writing behaviors, writes explicit judge.model in every behavior YAML, writes pinned assert-ai-version in the generated workflow.

Notes

• Assumes the reviewer replaces <LATEST_STABLE_AT_AUTHORING_TIME> in the skill examples with an actual pin when running the skill. That's the point — it's meant to be filled in per authoring.