Add first-class skipped node outcomes

[magix022]

Summary

• add ava.skip(reason, metadata=None) as a successful non-value node outcome
• propagate authored skips consistently through local, Ray, operator/protobuf, and TUI paths
• omit skipped fan-in slots, persistence writes, and lineage entries while keeping downstream dependencies complete
• define rerun behavior and document the public contract

Verification

• uv run ruff check src/ test/ — OK
• uv run pytest test/skipped_outcome_test.py test/operator_tests/test_operator.py::TestOperatorLifecycle::test_authored_skip_is_recorded_and_run_succeeds test/tui_test.py::TestLogTimestamps::test_log_panel_renders_authored_skip_reason_and_metadata -q — 8 passed
• uv run pytest test/dag_test.py test/workflow_execution_test.py test/workflow_data_passing_test.py test/rerun_test.py test/operator_tests/test_hooks.py test/operator_tests/test_operator.py test/operator_tests/test_grpc.py -m 'not ray and not tmux' -q — 160 passed, 28 deselected
• uv run pytest test/workflow_execution_test.py test/workflow_data_passing_test.py test/operator_tests/test_hooks.py -m ray -q — 8 passed, 13 deselected
• uv run pytest test/smoke_test.py test/example_smoke_test.py test/operator_example_discovery_test.py -v — 11 passed

Closes #13

[glesperance]

Bottom line

REQUEST CHANGES — skip compaction is not shared by provider matching and Python binding, so a skipped branch can hide a live Stream input. Three other public-boundary inconsistencies make accepted skip outcomes non-durable or misbound. Reviewed exact head a2e1cf2c4c61ba5cc59626a75f78617bd413fff2 against merge base c1248186551a3d56b57c62ae14f1eb46397f9466.

Findings

1. High — partial skipped fan-in selects the wrong Stream parent (src/avalanche/dag.py:1657, provider selection near :2671). Skip removal happens only during worker-side Python binding, after Stream has selected a positional parent from the unfiltered list. Reproduced with (skipped() & live_append()) >> consume(df=Stream(...)): the workflow returned None and bound Stream to ('skipped',) instead of the live AppendResult. Use one skip-compacted logical-slot plan for implicit Python and upstream-consuming providers, preserving worker-side Ray selection without driver payload fetches.

2. High — accepted metadata can make a successful skip impossible to read (src/avalanche/types.py:30, src/runtime/operator/convert.py:96, src/runtime/executor.py:77). ava.skip() accepts arbitrary mapping values, but operator serialization later requires JSON. Reproduced: Path and bytes metadata are accepted, then node_state_to_proto raises TypeError; integer keys change on round trip, empty reason becomes absent, and unbounded metadata rides the small Ray status channel. Validate and recursively clone a bounded, strict JSON-safe schema at skip() with a nonblank bounded reason.

3. Medium — skipped explicit positional-only inputs shift later values (src/avalanche/dag.py:1679). Removing the skipped element before call reconstruction compacts later explicit values into earlier slots. Independent audit reproduced consume(skipped(), live()) against consume(a="A", b="B", /) as ("LIVE", "B"). Preserve logical parameter identity, use the skipped slot's default when available, and reject skipped required positional-only slots clearly.

4. Medium — per-slot multi-return skips silently lose skip status (src/avalanche/types.py:45). Reproduced: a num_returns=2 node returning (ava.skip("left missing"), "RIGHT") emits success and returns (None, "RIGHT"); the reason is not durable. Issue #13 defines a node-level outcome. Reject nested or per-slot SkipOutcome values for v1, or explicitly design slot-level status before accepting them.

Execution-service behavior

Skipped tasks still finalize and produce receipts; receipt fan-in remains separate from payloads, and Ray observes small status refs rather than task payloads. The focused skip/operator/TUI/execution-services suite passed 28/28, including real Ray. Preserve these properties while fixing the four boundaries above.

Merge and evidence state

The live PR was CONFLICTING; merge-tree conflicts were in CHANGELOG.md and src/tui/widgets/log_panel.py. Resolve against current main while preserving virtualized rendering and authored-skip display, then rerun focused TUI/operator tests. Missing regressions: partial/all-skip Stream fan-in under Local and real Ray, positional-only skips, per-slot skips, and hostile or oversized metadata.