Resolve the three deferred TinyAgents §5 design decisions (recommend_tier, NotOwned, ToolScope/is_concurrency_safe)

[senamakel]

Summary

The TinyAgents port audit (§5 of docs/tinyagents-port-plan.md) surfaced three code-quality items that are genuine design decisions, not mechanical fixes. The behavior-preserving parts were landed in #4542, but each carries a real "implement vs defer" choice that was intentionally left open. This issue tracks resolving them.

Problem / Context

Each item currently reads as dead/misleading code but changing it alters behavior, a public type, or a trust boundary — so it needs an explicit decision rather than a drive-by change. #4542 added clarifying comments/annotations in place; this issue is where the actual decisions land (some are gated on later port phases).

Scope

Three independent decisions:

1. recommend_tier RAM scaling (src/openhuman/inference/presets.rs) — reads device RAM but always returns MVP_MAX_TIER; its test pins the non-scaling contract. Decide: implement RAM→tier scaling (and update recommend_tier_scales_with_ram), or formally keep the intentional MVP cap and drop the unused RAM read. fix(tinyagents-audit): execute §5 fix-in-place items — provider diagnostic, anchor coupling, dead-read clarifications #4542 documented it as an intentional cap.
2. SteerError::NotOwned RPC-path ownership (src/openhuman/agent_orchestration/subagent_control.rs) — the arm is unreachable on the trusted RPC steer_control path (no ownership check there); the agent-tool running_subagents::steer path does enforce ownership. Enforcing it on the RPC path would change the JSON-RPC payload shape (forbidden during the port) → deferred to Phase 4 (graph::orchestration::SteeringRegistry ownership consolidation).
3. ToolScope::AgentOnly vs All + is_concurrency_safe (src/openhuman/tools/traits.rs) — AgentOnly has three real constructors but no execution path gates on it (behaves like All); is_concurrency_safe maps to tinyagents ToolRuntime.idempotent, which the harness never reads for dispatch (serial tool loop). Reconcile with the crate tool model in Phase 2; adopting concurrency needs upstream parallel dispatch.

Acceptance criteria

• recommend_tier — decision made and implemented (scale or formally cap + strip param)
• NotOwned — RPC-path ownership resolved as part of Phase 4 SteeringRegistry, or explicitly documented as permanently trusted-RPC
• ToolScope::AgentOnly / is_concurrency_safe — reconciled with the tinyagents tool model in Phase 2 (gate AgentOnly or collapse to All; wire concurrency only behind upstream parallel dispatch)
• Diff coverage ≥ 80% — implementing PR(s) meet the changed-lines coverage gate

Related

• Port plan: docs/tinyagents-port-plan.md §5, §2
• Landed context: fix(tinyagents-audit): execute §5 fix-in-place items — provider diagnostic, anchor coupling, dead-read clarifications #4542 (behavior-preserving parts), feat(tinyagents): Phase 0/1 port cutover — pin v1.7.1, crate-back SchemaCleanr / model-context / observed stream #4539 (Phase 0/1 port)
• Tracker: TinyAgents migration (v0.58.7 → main) parity & bug-fix tracker — harness audit findings #4470 (TinyAgents migration parity)