feat(research): wire factor-profiles producer into the graph (un-orphan compute_and_write_factor_profiles)

[cipher813]

What

Splice a new compute_factor_profiles_node into graph/research_graph.py so the previously-orphaned producer scoring.factor_scoring.compute_and_write_factor_profiles (zero production callers, test-only → s3://alpha-engine-research/factors/ empty in prod) actually runs every Saturday SF research run.

Splice point + why

fetch_data → load_regime_substrate_node → macro_economist_node
  → compute_factor_profiles_node → compute_focus_list_node → dispatch_sectors_and_exit → … → score_aggregator

Spliced on the macro_economist_node → compute_focus_list_node edge (one re-route: that edge becomes macro → compute_factor_profiles_node + new compute_factor_profiles_node → compute_focus_list_node):

• The producer needs sector_map + run_date, both populated in fetch_data and not mutated by load_regime_substrate_node / macro_economist_node.
• It must write factors/profiles/{run_date}/by_ticker.json + latest.json before both consumers' existing read_factor_profiles_from_s3(): compute_focus_list_node (:1198) and score_aggregator (:1322, downstream of the dispatch that hangs off compute_focus_list_node). This one splice satisfies both.
• The fetch_data → load_regime_substrate_node edge was not used because the Stage-C serial chain is a pinned topology invariant (tests/test_regime_stage_b_graph_topology.py). The macro→focus-list edge is the cleanest splice preserving every pinned edge and the regime / macro / focus-list / dispatch chain + conditional edges.

No SF/infra change required — it is a plain in-graph serial node.

Graceful-degrade

The node wraps the producer so any failure (missing run_date, missing/short features/{run_date}/*.parquet, S3 error, compute exception) is caught, logged flow-doctor-visibly (logger.warning/logger.error), and the node returns cleanly ({"factor_profiles_written": False, "factor_profiles_s3_key": ""}). The graph continues; the consumers degrade exactly as they do today when the substrate is absent (they already if not factor_profiles: skip) — no worse than the prior orphaned state. The weekly research run is never hard-failed on this new dependency. Profiles are not threaded through state — consumers read from S3 by design; only a small observability delta flows.

Behavior-safety

No flag is flipped. config.FACTOR_BLEND_ENABLED and config.FOCUS_LIST_GATING_ENABLED stay default-false. This is substrate-only: it makes the factor substrate exist/ready and lets the focus-list shadow audit populate scanner_evaluations.focus_*. No scoring or agent behavior changes.

Closes / unblocks

• Closes ROADMAP P1 "Wire the orphaned factor-profiles producer into the Saturday SF".
• Unblocks the FOCUS_LIST P0's real gate — its shadow audit now sees a populated factor substrate each run.

Tests

• tests/test_factor_profiles_node.py (new): (a) node calls compute_and_write_factor_profiles with the state's run_date + sector_map and returns the observability delta on success; (b) producer exception / missing run_date → node logs + returns cleanly, no raise; (c) static-AST graph-wiring assertions (mirroring test_regime_stage_b_graph_topology.py) that the node is registered, runs after fetch_data via the macro chain, and strictly before compute_focus_list_node AND score_aggregator, without altering the sector dispatch.
• tests/test_dry_run.py: fixed a pre-existing order-dependent test-isolation bug surfaced (not caused) by the new test file. TestGraphModuleGuard.test_skips_late_bound_patches_when_graph_absent and TestInstallRestore setup/teardown evicted/replaced the real sys.modules["graph.research_graph"] without restoring it; a later re-import created a second module object so other test modules' collection-time-bound _build_signals_payload no longer saw their monkeypatch.setattr("graph.research_graph.<FLAG>", …) (the exact leak the test_regime_stage_b_graph_topology.py docstring documents and previously only "sidestepped" via filename ordering). Both sites now snapshot + restore the real module.

Full suite: 1366 passed (~/Development/alpha-engine-research/.venv/bin/python -m pytest -q).

🤖 Generated with Claude Code