fix(kernel): scope collect step to preceding fan-out outputs

[andyst-dev]

Fixes #1253

Summary

The collect step joins the whole run-wide all_outputs buffer instead of only the outputs of the immediately preceding fan-out group. A workflow shaped sequential → fan_out → fan_out → collect therefore feeds the earlier sequential step's output into the collected payload, and multi-phase workflows keep carrying stale pre-fan-out context across later collect steps.

docs/workflows.md specifies: "The collect step gathers all outputs from the preceding fan-out group" — the implementation did not match that contract.

Changes

• execute_run now keeps a fan_out_start marker, reset to all_outputs.len() at the beginning of each fan-out group.
• The collect step joins all_outputs[fan_out_start..] instead of the full buffer, so it merges only the outputs produced by the immediately preceding fan-out group.
• Added a regression test covering sequential → fan_out → fan_out → collect, asserting the collected payload contains both branch outputs and not the earlier sequential output.

Behavior for workflows without a preceding fan-out group is unchanged (the marker defaults to the start of the buffer).

Testing

• cargo test -p openfang-kernel — all tests pass (workflow tests: 18 passed)
• cargo fmt -p openfang-kernel -- --check — clean
• cargo clippy -p openfang-kernel --all-targets -- -D warnings — only a pre-existing clippy::question_mark warning in crates/openfang-channels/src/feishu.rs:1052, present on main and unrelated to this change; the changed file produces no warnings

Security

• No new unsafe code
• No secrets or API keys in diff
• User input validated at boundaries