feat(analyzer): P4 scoring methodology -- one normalization regime, personas, de-circularized health

What

P4 scoring methodology rework, landed as one coherent change (the pieces are coupled: the normalization scale change ripples into the archetype thresholds and the health score).

• One normalization regime for all six domains: winsorize the upper tail, log1p re-express, then min-max rescale (replaces the mixed percentile-vs-threshold split). Follows the OECD composite-indicators method.
• Archetypes reframed as interpretive personas (a design heuristic), not a derived clustering.
• Health de-circularized: removed the archetype-diversity input that fed on the classifier output.

Result (validated on a 300-user seeded instance)

The old demo-fit thresholds put 46% of users in the "developing" fallback with zero platform engineers. The better-shaped scale corrects this without gaming thresholds:

• developing 138 (46%) to 61 (20%); all eight personas now represented (platform engineer 0 to 20).
• No domain collapse: source control median 0.444 to 0.690.
• Health 43 to 45, no longer mechanically driven by the classifier.
• Output is deterministic (re-run is byte identical).

Grounded in SPACE, the OECD composite-indicators handbook, Box-Cox/Tukey re-expression, and persona-design literature.