[senamakel]

Adds a generic TinyAgents turn model set that builds the primary model, workload route models, and a distinct summarizer through borrowed host factory and gate ports. It preserves requested aliases separately from resolved routes, supports host overrides and per-purpose options, skips denied/unavailable workload routes, and provides subprocess builders for existing Claude providers.

The returned set can produce an owned resolver and ordered primary/workload registrations from the same host-decorated model instances. This lets OpenHuman retain its existing budget wrappers and turn-plan error policy without reconstructing provider selection. Construction and subprocess tests use fakes; they do not probe a real CLI or network. The existing Claude subprocess error tests now use a checked-in executable fixture with per-test response files, avoiding runtime executable creation while retaining real spawn and output assertions.

Regression evidence: prototype runtime tests failed on the default workload role alias and primary-alias skip behavior before those fixes. Later behavior tests use the actual resolver, borrowed factory/gate fixtures, and host wrapper semantics. An independent full-suite run then exposed ETXTBSY while spawning a runtime-written fake Claude executable; the fixture now uses a static executable and the full package suite passes.

Validation: cargo test -p tinyagents-harness (2,994 unit tests and 44 doctests passed, 3 ignored), 31 Claude driver tests, focused model-factory tests with default and no-default features, cargo check -p tinyagents-harness, strict all-target clippy, and cargo fmt --all -- --check.

Refs #73
Part of tinyhumansai/openhuman#8092