v0.1.6: operational validation pass

Closes the remaining gaps from the v0.1.5 evaluation: untested resume-protocol heuristic, no end-to-end smoke test, oversized description for Cursor's per-rule limit. Adds CI integration of the smoke test.

Added

• scripts/dogfood-smoke.sh: 7-test operational smoke test covering drift-detection (happy + synthetic), next-sub-step heuristic, artifact-path contract reachability for all eleven .<tier>-ready/ paths, AGENTS.md emit-respect (existing + absent cases), critical-finding gate. 7/7 pass against a synthetic mid-arc project. Wired into CI on push and PR.
• .cursorrules (1,754 chars; under Cursor's 2KB recommendation): slim trigger-surface adapter for Cursor users. Points at the full SKILL.md. The full description (2,558 chars) is fine for Claude Code, Codex, Antigravity, Pi, OpenClaw.

Changed

• SKILL.md resume-protocol shell snippet: walks tiers in dependency order (1.1 -> 1.2 -> ... -> 3.4), treats skipped as complete, resolves next-sub-step against dependency-order traversal not file-order pattern matching. Smoke test verifies it returns 1.2 correctly when PRD is imported and ARCH is in-flight (the bug v0.1.5 acknowledged).

Validated

Check	Result
Description size	2,558 chars / ~639 tokens; under 4KB practical ceiling for Claude Code / Codex / Antigravity
Cursor adapter size	1,754 chars; under 2KB Cursor recommendation
Trigger collision (static)	63 unique domain-specific triggers; no generic single-word triggers; low collision risk
Operational smoke test	7/7 pass

Still untested (requires live harnesses; out of scope for build script)

• Functional agent execution end-to-end on a real or fictional product (a session, not a script).
• Live trigger routing on Claude Code / Codex / Cursor / Windsurf (empirical measurement).
• Real cache cost across the eight compatible_with harnesses.

Why a patch

Refines an inline heuristic; adds a smoke test + adapter; no discipline change. arc-ready remains faithful consolidation.