Phase 1: Routing agent — per-connected-component deterministic/agentic decision

[matthewmoorcroft]

Summary: Turn motifs and insights into a user-approved deterministic/agentic conversion plan at connected-component granularity; this Phase 1 routing step is not built.

Depends on: #64, #25/#82, and ADF control lineage from #80/#81 after their main re-land.

Blocks: #87 and supplies decisions later consumed by #85.

Current status

This is an open issue with no routing implementation yet.

The required inputs exist only on staging or in follow-on work:

• discovery_inventory.py:build_source_inventory emits per-pipeline control lineage on integration/discovery via Map ADF onto the shared source AST + source-agnostic inventory emitter (#62a) #79-Emit discovery-AST lineage into inventory.json (#62) #81.
• discovery_insights.py:enrich_inventory adds agent-authored insights on integration/discovery via Agentic insights on the unified inventory (discover phase) (#25) #82.
• Motifs are currently only counted during discovery and detected/collapsed during conversion by sources/adf/loader.py:build_profile_rows, motifs/detector.py:detect_motifs, and motifs/collapser.py:collapse_motifs; additive motif capability annotations are still Phase 1: Additive motifs in inventory as the routing capability signal #64.
• Existing adapter decisions in src/flowx/adapter/operations.py modify per-activity translation choices after IR exists; there is no factory-level connected-component route plan.

What needs to change

Add an agent-authored, library-validated conversion-plan artifact that:

• reads the enhanced inventory (motifs as capability signal plus insights);
• recommends deterministic 1:1 conversion wherever the engine is capable and agentic conversion where it is not;
• presents the recommendation to the user and records the user's decision;
• computes connected components over cross-pipeline ControlEdges and requires one coherent route per component, so caller/callee references are never split across incompatible routes;
• remains descriptive only: Phase 1: Routing agent — per-connected-component deterministic/agentic decision #77 records routing; Phase 1: In-engine agentic escape hatch (passthrough + post-convert IR alteration) #87 performs Phase 1 post-convert IR alteration and ADF: map discovery AST -> IR (carry lineage through conversion) #85 later consumes the same decision during Phase 2 conversion.

No decision means today's behavior. Routing is opt-in.

How to approach

Add a source-neutral plan model/validator adjacent to models/insights.py and discovery_insights.py, reusing the latter's inventory fingerprint and pipeline/control-edge lookup patterns. Build weak/undirected connected components from each ControlEdge source_workflow/target_workflow pair, retaining unresolved edges as validation findings rather than silently severing them.

Expose an explicit author/validate/record step through adapter/main.py and mcp/server.py. Store the deterministic recommendation, user choice, component membership, evidence (motifs/insights), and inventory fingerprint. Do not mutate TranslationReport or Pipeline here.

Acceptance/verification:

• Component tests cover chains, branches, cycles, isolated pipelines, unresolved callees, and multiple invocation edges.
• Validation rejects partial or conflicting decisions inside one component and unknown pipeline names.
• Recommendations default to deterministic when motif capability covers the component; agentic recommendations cite uncovered gaps/insights.
• The plan is fingerprint-bound, deterministic/idempotent to serialize, and fully user-reviewable.
• With no plan, existing ADF and Airflow conversion/package outputs are byte-equivalent to today.
• Airflow's current path is untouched; Pete may adopt the source-neutral artifact later.

This is additive, opt-in routing metadata only.