pa-01 — Analysis

A design-review treatment of the service-graph analyzer and the decomposition decisions it informs.

What the analyzer must get right

Sound cycle detection. Report exactly the strongly-connected components of size > 1 (plus self-loops); a clean DAG reports nothing.
Correct blast radius. Transitive dependents (reverse edges), not dependencies; exclude the service itself; deterministic (sorted).
Honest coupling metrics. Fan-in/out are direct-edge counts; blast-radius is the transitive version of fan-in.
Enforceable layering rules. A rule is a (from-layer, to-layer) prohibition checked over every edge.

Edge direction = "depends on." A→B means A needs B, so failure propagates against the edges (B's failure breaks A). Blast radius therefore walks edges in reverse. Getting this direction right is the whole correctness of the tool.
Tarjan over naive DFS-coloring. Tarjan returns the actual SCCs (the cyclic groups), which is what you report to engineers — not just a boolean "has cycle." Deterministic via sorted iteration.
Layers as free-form tags. Rather than hard-coding a layer model, a service carries a tag and rules are data, so the same engine enforces "domain↛infra," "no upward deps," or team-ownership rules. This is the hook for pa-10 fitness functions.

The graph is only as true as its source. Edges from a wiki are fiction; derive them from import graphs, call traces, or a service registry. Garbage in, confident-but-wrong out.
Cycles aren't always fixable by splitting. Sometimes two services genuinely co-evolve and should be merged back into one. The tool flags the smell; the architect decides split vs merge vs async-break.
Low coupling vs too many services. Driving fan-out to zero by splitting more creates nano-services (more hops, more contracts). The metric guides; it doesn't dictate.
Static structure misses temporal coupling. Two services with no edge can still be coupled if they must change together for a feature. Pair the graph with change-coupling analysis (files/services that commit together).

Auto-derived graphs from tracing/mesh + git history (change coupling).
Weighted edges (call volume, latency) for traffic-weighted blast radius.
Contract-compatibility gates (pa-02) wired into the same CI as the cycle/layering fitness functions (pa-10).
Ownership/team metadata (Conway's-Law analysis: do boundaries match teams?).