feat(init): teach the next step and mark detected targets

[Roger-luo]

What

Dogfooding ion init on a real project (bloqade-website) surfaced two progressive-prompting gaps on the init journey. Both are fixed here, test-first.

1. init ended with no next step (both channels)

The old output stopped at Created Ion.toml with N target(s) — a dead end for a human and an agent. Now a single, state-aware next-step line renders from one source of truth to both channels:

• fresh project → ion add <source> / ion search <query>
• skills already declared → ion add

The --json envelope gained a data.next array:

"next": ["ion add <source>", "ion search <query>"]

2. No-TTY init told the agent more than the human (channel mismatch)

--json exposed a per-target "detected": true flag, but the human plain-text list showed nothing and hard-coded --target claude in the recovery example. Now the human list marks (detected) and prefers a detected target in the example:

Available targets:
  claude -> .claude/skills (detected)
  cursor -> .cursor/skills
  windsurf -> .windsurf/skills
Re-run with --target <name> to select targets (e.g. --target claude).

Changes

• src/commands/init.rs — next-step helpers (next_steps_after_init) + (detected) marking
• src/builtin_skill.rs — expose SKILL_NAME so the next-step counter excludes the Ion-managed builtin
• tests/json_integration.rs, tests/scenario_tests.rs — three failing-first tests pinning the guidance in both channels
• skills/dogfooding-ion/SKILL.md — logged the second-pass findings

Notes

ion-cli = { type = "local" } is intentional (skipped by fetch/validate/update, refreshed via refresh_global) — documented, not changed.

Test plan

• cargo fmt --all ✅
• cargo clippy --all-targets --all-features -- -D warnings ✅
• cargo nextest run — 800/800 ✅

🤖 Generated with Claude Code