Unified compaction, thinking as ability, find_tools v2

The compaction redesign collapses multiple entry points into a single _compact() method, retiring the Providers singleton and dead helpers. The model now writes a transcript row with role=‘compaction’ whose id serves as the watermark, replacing the old tool_calls mechanism. A LIMIT 20 bug on the since_id branch was removed, fixing the proactive >50-turn compaction check.

Downstream adjustments ensure compaction rows do not leak into user-facing views. Chat history, episode extraction counts, and window queries now exclude role=‘compaction’ rows. Observability tool-usage stats filter the new internal trail_compaction meta-tool name. Additionally, the

extraction step was restored so only the dense summary body persists, discarding the scaffold.

Tests, docstrings, architecture docs, and spec documents were reconciled to match the new model. Tests now seed real transcript rows for compaction assertions and patch the sanctioned LLM boundary. Documentation reflects the unified _compact() flow, watermark mechanism, and the corrected LIMIT and summary-extraction decisions.

The thinking ability was refactored into an internal, never-discoverable ability registered via file-scan but excluded from find_tools indexing. It is dispatched at turn 0 through ToolDispatcher when thinking_level is ‘high’, replacing the previous ThreadPool exploration hack. A dead attribute wiring bug that prevented high-thinking dispatch was fixed, and the SHA sidecar was corrected to match the 35-ability db index.

find_tools v2 introduces a split select/query interface, drops the limit parameter, and applies a relevance score floor to filter junk results from the query path. Dead code (store_batch, get_find_tools_results) was removed.

Infrastructure: Superpowers dev plans were untracked and internal doc references stripped; skill-association now uses its own SkillAssociationConfig loop; TrailHandoverConfig was added for act-trail handover compaction.

Compaction unified under _compact(); Providers singleton retired; LIMIT 20 bug fixed.
Compaction rows (role=‘compaction’) excluded from chat view, episode extraction, and tool-usage stats; only extracted
persists.
Thinking ability is now internal, dispatched at turn 0 via ToolDispatcher; dead high-thinking gate fixed.
find_tools v2 splits into select/query with relevance floor, drops limit, removes legacy dead code.
All compaction tests, docs, and specs reconciled to the transcript-watermark model.