fix(dw-pim-completeness): per-language scoring with the flag on, MCP calculation and variant-master traps (5.8.8)

[Justinvolved]

Folds four measured learnings into the completeness family:

• Completeness is per language once the "Completeness feature" flag is on. The "no per-language dimension" claim is scoped to the flag-off path. With the flag on, the index builder writes CompletionRule|<id> per index document in that document's language; recipe (flag, SkipCompletionRules=false, group and shop completion languages, full rebuild, LanguageID + empty VariantID + CompletionRule|<id> < 100), copy-fields-only rule, and the admin per-language badge. The "buggy beta" warning is narrowed to DW 10.28. Measured on DW 10.29.6.
• What calculate_product_completeness_for_products returns. Family-aggregate value per master, master-row-only fieldStatuses, languageIds ignored; the default-image index fields are empty, only ImageSmall drives a missing-image query.
• Masters with variants and ProductNumber. Combine stores the master number as '' and the master edit screen has no Number field, so keep ProductNumber out of rules (DW 10.29.7, reported upstream as PIM: Combine as variants clears the master's product number, so a completion rule with Number cannot reach 100 % on a new family DynamicWeb#735). Corrects the earlier reading: variant-editable category fields are checked per variant. Combine side effects and two variant-field traps added to structural-model.md; excludeVariants note in dw-pim-workflow.
• assign_completion_rules_to_shops persists only the last request of a batch (MCP 0.6.0-BETA). One call per shop, read back; row in tool-surface-gaps.md.

Version 5.8.8. validate-skills.py exits 0 and build-manifest.mjs --check is clean.