Skills from fixed errors, consolidation as proposals, memory linked to the knowledge graph #1157

Follow-up to #1149, #1152 and #1154. Three gaps in how the workspace learns and connects what it knows:

1. Skills from fixed errors. The most reliable learning signal is not an optional _intent parameter but a call that failed and then worked once the arguments changed (a date format, an id instead of a name). Today nothing turns that into a rule.
2. Consolidation rewrites active skills directly. "Consolidate with AI" deletes the applied skills of a scope and writes the merged ones, live and without review, with new ids. That loses history and breaks anyone who references the old ids.
3. Memory and the knowledge graph are separate. A fact about customers is not offered when an AI client asks the graph how to get a customer.

Proposal

Skills from fixed errors

• Stamp each tool call with whether its caller contributes to learning at that moment (tool_invocations.learnable, false by default and for everyone who does not contribute; erased together with intents when a member is switched off).
• A pass over the learnable calls of the last 30 days finds a failed call followed within 10 minutes by a successful call of the same tool by the same member with different arguments. It records only the argument names and the format of the values (for example 99/99/9999 → 9999-99-99), never the values.
• A pattern seen at least twice becomes a pending skill on the tool's connector (source errors), with how often it happened. No AI model is needed, so it works on every instance. It runs from the AI Skills page and with skill generation; dismissed proposals are not proposed again.

Consolidation as proposals

• Consolidation creates pending merged skills (source consolidation), each recording which applied skills it replaces. Skills are only merged within the same folder, and a merged skill keeps that folder.
• Applying a merged skill dismisses the skills it replaces (kept, with their history); dismissing it changes nothing.

Memory linked to the knowledge graph

• A memory item can be about an entity of the graph (about, e.g. customer). memory_save and memory_search accept it.
• kg_how_to_obtain adds the facts about the matched entities that the caller may read (memory access, folders), labelled as data.

Migration

tool_invocations.learnable (boolean, default false: a catalog-only change), kg_skill_suggestions.replaces (JSON, default []), memory_items.about (nullable) with an index. No data is removed.