D1197 — facet_sprawl proposes a closed tags vocabulary that map_update and kb_repair apply

Decision. The facet_sprawl observation carries a deterministic Proposal (field_values.tags: values, mapping, drop) built from clusters of tag spellings, and invalid_field_value gains a list-aware fix, list_value (From to To, empty To removes), reported per bad item. A removal is applied only by an explicit kb_repair; the write path and the background repair apply renames only.

Why. The observation only listed the most used tags, so a delegated doctor session could not close it and the item came back every session; missing_value_contract already showed the shape that works (a ready proposal, map_update, kb_repair). The tags field is a list, which the scalar set_value fix could not repair. Costs: once declared, the vocabulary is closed, so a new tag needs a map_update first.

Alternatives rejected. Edit distance or co-occurrence to merge more spellings: not deterministic enough to explain to an operator, and a wrong merge rewrites pages. Storing the drop list in the contract: a new contract key for a one-shot action. Applying removals unattended like other AutoRepairSafe fixes: it would silently erase a tag an agent writes before the map_update that declares it, turning "declare first" into data loss. A cluster threshold of 1: every singleton would enter the vocabulary and the sprawl would stay.

Consequences. Fix has a from field and Proposal a drop list; the keep predicate unattendedFix (internal/mcpserver/tools_repair.go) must stay on every unattended path. map_update replaces every field_values.* and value_synonyms.* key, so the doctor merges with the existing ones. value_synonyms is map-wide: a vocabulary tag that belongs to a built-in status family can leave an item without a fix, never with a wrong one.