Solver/CLI behavior items from bug sweep (self-loop iteration, observation validation, dead --mode)

[adamjohnwright]

From a bug sweep of the live parse/solve/export path. The three safe, non-benchmark-affecting fixes from the same sweep were already applied (SCC convergence residual now reports the worst loop instead of the last; REACTOME_CACHE guarded with a lock; multipart upload filename sanitized). The items below change behavior/output and need a decision + likely a benchmark check.

1. Self-loop nodes evaluated once instead of iterated to a fixed point (major)

src/solvers/steady_state.jl solve_scc_ordered! classifies a component as acyclic purely by node count (comp_size[c] == 1). A node with a genuine self-edge (a reaction whose target is also one of its own activators/inhibitors, i.e. parent_uuid == child_uuid) is a real 1-node cycle but takes the acyclic single-eval branch, reading its own pre-update value once instead of relaxing.

• Fix: route size-1 components that have an intra-component input edge through the iterated (loop) branch.
• Impact bounded by how many literal parent==child reactions the decomposed networks contain — needs a count + benchmark check before changing.

2. Observation condition column silently ignored (major)

cli/deltasignal.jl keys observations by node UUID only; condition is never read. Any observations file where a node appears under more than one condition collapses last-write-wins with no error, and per-condition solving isn't possible. Decide: group/solve per condition, or explicitly error on duplicate node keys.

3. No range/scale validation on observation values (major)

The CLI accepts any value with no clamping, while the documented contract is 0–100 (internal = ui/100). The bundled examples/sample_observations.csv is actually log2-fold-change data (negatives, values <1), so the documented example "succeeds" while producing meaningless clamped output.

• Fix: validate/clamp observation values to [0,100] in the CLI (and warn), and correct the sample file to the 0–100 scale.

4. Dead --mode solve flag (minor)

cli/deltasignal.jl defines --mode but never reads it; the method is hardcoded to "penalty". Remove the flag or thread it into SteadyStateParams.

Also noted (minor, no action assumed)

• --output (0–100) vs --output-pathway (0–1) scale mismatch in the same solve.
• export --format pathway-browser doesn't aggregate despite the docs listing aggregation methods under export; export_pathway_browser_overlay in reactome_mapper.jl is dead code.
• parse_set_mappings docstring describes a 3-column format but the code requires the 4-column long form (sample matches the code).

🤖 Generated with Claude Code

[adamjohnwright]

Item 1 (self-loop) resolved as a non-issue on the live path.

Investigation: the generated logic_network.csv output is entirely UUIDs (checked: 0 stable-id sources, 231,630 UUID sources in one top-level network; 0 source==target edges across 2.38M edges in all 29 generated catalog networks). deltasignal groups edges by child to form its (target, sources) reactions, so no self-loop can reach the size-1 acyclic misclassification in solve_scc_ordered!. The inh == target_idx guard in reaction_model.jl:466 is defensive-only and never fires on real data; positional decomposition deliberately severs autoregulation into feed-forward (reaction_model.jl:511-514).

The only path that produces source==target self-loops is bench/benchmark_selective_joint.py, which collapses UUIDs → stable_ids (merging the many positional instances of one recycling reaction onto a single stId). That's a research/benchmark harness, not the product path.

Fix applied there: drop self-loop edges during the collapse (if s == t: continue, with a dropped-count print), so they can't take the wrong solver branch. No change to solve_scc_ordered! — the hot loop stays untouched and no live benchmark number moves.

Closing the self-loop portion. Remaining open items on this issue: observation condition ignored, observation scale validation, dead --mode flag.