fix(epic): default --discover to parent-child (stop blocks-edge sibling-epic leak)

[ryanholtschneider2]

What

Changes the default --discover mode for po run epic and po run epic-wts (and the epic_wts wrapper) from both to parent-child.

Fixes po-formulas-software-dev-e9s.

Why

both/deps walk blocks edges up to collect children, so they widen the discovered set across epics. When EpicB blocks-depends on EpicA, dispatching EpicA discovered EpicB + all of EpicB's children and ran them in parallel — inverting the dependency. The CDR-A/CDR-B incident in the issue discovered 15 nodes for a 7-child epic.

blocks is the right edge for the dispatcher to order by, but the wrong edge to widen by. parent-child discovery collects children via parent-child edges only; blocks edges are still honoured for topo-ordering within the discovered set (list_subgraph always builds the blocks sub-DAG regardless of traverse). Operators who want the old union opt in with --discover both.

Changes

• parent/po_formulas/epic.py + wts/po_formulas_wts/epic.py: default discover="parent-child"; new _discover_children helper routing parent-child → list_subgraph(traverse=("parent-child",)); accept parent-child as a valid discover value alongside core ids/deps/both.
• wts/po_formulas_wts/epic_wts.py: default discover="parent-child" (forwarded to epic_run).
• parent/tests/test_epic_discover_flags.py: default-is-parent-child test, explicit parent-child test, and both-opts-into-union test (replaces the old default-is-both test).
• parent/README.md + wts/README.md: discovery-modes table + default updated, with a callout explaining the leak.

Acceptance (issue)

• po run epic --epic-id <X> discovers only parent-child children by default.
• Operators wanting the blocks-aware union opt in via --discover both.
• Docs updated.

Verification

• test_epic_discover_flags.py (incl. 3 new/updated tests) + test_epic_legacy_dot_suffix.py + test_epic_wts.py: green.
• Close-the-loop: reproduced the CDR-A/CDR-B incident against the real core list_subgraph — discover=both returns [A1,A2,B1,B2,EpicB] (the leak); discover=parent-child returns [A1,A2] only. Repro + output saved under the run's review-artifacts/.

Pre-existing failures (NOT from this change)

Two failures exist on a clean origin/main full-suite run in this degraded dev env (the editable prefect-orchestration resolves to a stale stub; the formula packs aren't EP-installed in the shared venv):

• parent/tests/test_software_dev_pack_path_metadata.py::test_bd_metadata_overrides_rig_path_default — test-ordering pollution; passes standalone, fails in full suite on clean main too.
• wts/tests/test_pack_skeleton.py (2 tests) — assert cross-pack entry-point coexistence; the isolated single-pack venv has no parent-pack po.formulas entry points.

git diff origin/main confirms software_dev.py and the affected test files are byte-identical to main, so neither failure is attributable to this change.