Add unwrapped dict task inputs

[Volv-G]

Summary

Adds a first-class @task(unwrap=...) API for Python-authored tasks that need to receive a caller-provided dict as explicit component inputs instead of one JsonObject input.

@task(image="python:3.12", unwrap="run_data")
def combine(run_data: dict[str, str]) -> str:
    ...

combine(run_data={"shop": shop_task.Output, "catalog": catalog_task.Output})

The compiler now flattens the traced call-site dict into synthetic inputs such as run_data__shop and run_data__catalog, so normal argument/edge emission can wire each dict value independently.

Mechanism

• Extends @task(...) with unwrap="param" and unwrap=["param", ...].
• At trace time, flattens marked dict arguments into explicit task arguments (param__key) and records the exact key order per task call site.
• Persists a hydrate-safe local_from_python.unwrapped_inputs schema in the generated components sidecar:
  • original parameter name;
  • synthetic input prefix/name;
  • caller-provided key list/order;
  • inferred Tangle type from dict[..., V];
  • optional flag.
• Uses the same persisted schema in component generation during both compile and hydrate, so regenerated component inputs are deterministic without access to the original Python call site.
• Re-wraps flattened CLI args back into the original dict inside the generated runtime Python wrapper before calling the user function.
• Adds schema-hashed sidecar fragments for unwrapped tasks (task-name--<hash>) so the same function can be called with different dict key sets without component schema collisions.
• Preserves function-like introspection attrs across CallableRef.named()/bind()/with_annotations() so schema generation works for named task call sites.

Tests

Adds coverage for:

• @task(unwrap=...) API normalization/validation.
• Type inference from dict[str, V] into flattened input types.
• Generated runtime source re-wrapping flattened inputs into the original dict.
• End-to-end component generation of flattened inputs.
• Pipeline compile sidecar emission with local_from_python.unwrapped_inputs.
• Compile-time flattening of task arguments and task-output edges.
• Schema-hashed fragments for different key sets of the same function.
• Hydrate forwarding of unwrapped_inputs for compile↔hydrate round-trip determinism.
• Edge cases: empty dicts, non-dict annotations, invalid names/keys, and argument collisions.

Version

Bumps tangle-cli from 0.1.6 to 0.1.7 so this can be tagged/released after merge.

Design reference

Implements the design tracked in the PiForge project wiki page design-dict-unwrapping-task-inputs.

Local validation

• uv lock --check
• uv run ruff check packages/tangle-cli/src/tangle_cli/component_from_func.py packages/tangle-cli/src/tangle_cli/component_generator.py packages/tangle-cli/src/tangle_cli/pipeline_compiler.py packages/tangle-cli/src/tangle_cli/pipeline_hydrator.py packages/tangle-cli/src/tangle_cli/python_pipeline/ref.py packages/tangle-cli/src/tangle_cli/python_pipeline/task.py packages/tangle-cli/src/tangle_cli/python_pipeline/graph.py tests/test_component_from_func.py tests/test_pipeline_compiler.py tests/test_pipelines_cli.py tests/test_packaging.py
• uv run pytest → 805 passed, 4 warnings