Code review 2026-10-02: fail loudly instead of shipping a silently wrong bundle (default output unchanged)

[adamjohnwright]

These are the pipeline-hygiene findings from the 2026-10-02 code review. Default output is unchanged: GPVI, NODAL and RAF regenerated from this branch match build 20260928-1110_06ccb63 exactly, edge for edge by stable id, and every bundle file has the same row count.

Finding	Before	After
F2 export failures	Every export after logic_network.csv was logged and skipped. That included the orphan guard's deliberate ValueError, so containment.csv, cofactors.csv and node_resolution.csv were missing, the pathway counted as built, and catalog.sh moved current. Regenerating into an existing directory left stale files behind.	Raises, so the build is incomplete. Regeneration clears the bundle files first.
F3 LNG_* parsing	LNG_HANDOFF_EDGES=false turned harmful edges on. LNG_BOUNDARY_HIERARCHY=true turned the adopted hierarchy off. Typos were never read.	src/env_flags.py: a switch takes exactly 0 or 1, and an unknown LNG_* name stops the run at startup.
F5 modifier sets	One failed lookup cached a 6-set fallback for the whole run.	Raises; nothing is cached on failure.
F11 swallowed lookups	The Ub lookup dropped every ubiquitin depletion edge. Containment recorded "contains only itself". node_resolution dropped every set row.	Each raises.
F13 docs	LNG_DIAGRAM_BRIDGE was listed with default 1, and the DS defaults listed were stale.	Corrected. The DS defaults now point to the deltasignal CLAUDE.md.

Tests: every guard is exercised at runtime (test_export_failfast.py drives generate_pathway_file; test_env_flags.py; test_lookup_failures.py), and the old code fails them. Two source-text assertions were replaced by checks on the actual defaults. Default tier: 336 passed.

Not in this PR: the findings that change networks, F7, F1, F10, F9 and F8:

• a set component of a root complex becomes required AND inputs;
• set members leak between inputs;
• set-containing complexes are dissolved;
• variant pairing crosses members;
• the E3 ligase is treated as its own substrate.

F7 is verified in Neo4j. For example, "VAV1 Rho/Rac effectors:GDP" is GDP plus a CandidateSet of 5 GTPases, but the network requires all 5. These change benchmark numbers, so they go through a deltasignal spec with pre-registered arms.

🤖 Generated with Claude Code