[roodboi]

Problem

A HACK-1163 acceptance run on the M3 (next at ccf526c) found that native up refuses with graph_readiness when the compose file has a service in a profile that is not selected.

Cause: readiness() in src/backends/native-project-start.ts emitted --ready for every compose service. src/backends/native-project-input.ts does not filter by profile, which is correct, because the review needs the whole file. The runtime then skips inactive services in Graph::from_plan (packages/runtime-core/src/project/execution.rs) and refuses with "Readiness includes inactive or unknown services."

Change

• readiness() is now the exported nativeReadinessArgs({ specs, plan, initializers, routed }). It emits goals only for services that the reviewed plan (review.report.plan.services[name].active === true) marks active.
• depends_on edges declared by inactive services are ignored. Before, an inactive service with depends_on: { web: { condition: service_healthy } } could raise an active web to healthy. That fails the same way when web has no healthcheck.
• The runtime refusal is unchanged. It is a correct guard against goals the plan does not schedule.

Bridge pool capacity (native-project-routing.ts): kept as is

nativeBridgeCapacity counts routed services in inactive profiles on purpose. I left that unchanged and added a comment that gives the reason:

• It runs before runtime up and before the profile-aware review, so profile activation has not been decided yet. Duplicating compose profile and dependency activation in TypeScript would create a second source of truth.
• reviewedNativeRoutes refuses when the number of active routes is more than the capacity. Counting every declared route guarantees the floor is never too small, whatever the selected profiles are.
• The cost of over-counting is a few idle bridge sockets in the pool (capped at 32). That is harmless, unlike an extra readiness goal, which the runtime refuses.

Validation

• bun test tests/native-project-start.test.ts tests/native-project-routing.test.ts tests/integration-ownership.test.ts: 74 pass, 0 fail.
  • New unit test: inactive services (routed and unrouted) produce no goals, active ones do, services missing from the plan are excluded, and a missing plan produces no goals.
  • New startup tests through startNativeProject with injected runners (no Docker, OrbStack or real HOME): without a profile, only web=started is emitted, there is no admin=auto route slot, and bridge capacity is still 1. With profiles: ["debug"], the review receives debug and the goals become admin=healthy, web=healthy and worker=started, with the admin=auto route slot.
  • Negative control: forcing the active check to true (the old behaviour) makes the unit test and the no-profile startup test fail.
• Rust: the new execution_dependency_compiler_refuses_readiness_for_inactive_or_unknown_services contract test proves Graph::from_plan still refuses goals for an inactive profile service (tools) and for an unknown service (ghost) with graph_readiness, and accepts the active-only set.
• RUSTUP_TOOLCHAIN=1.97.1 bun run check:local (fmt and clippy -D warnings) and bun run test:local: all green (985 pass in the main suite, 0 failures).
• bun run --cwd packages/cli typecheck: 0 errors. bun run --cwd packages/cli check: 0 errors, and the existing warnings are not in the changed files.

Gaps

• I did not re-run native up on the M3 against a real project with an unused profile. This PR proves the argument construction and the runtime guard in isolation, not end-to-end runtime behaviour.
• I did not run the full bun run test or hosted CI locally.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}