ACT Visual Feedback & System Refinements

Inline tool start/end events and UI feedback are now available for ACT loops, showing progress via inline pills. Backend now mints stable call_id for tools and emits start/end events with success status. In the frontend, tool pills nest within the narration bubble or pending form and enforce a minimum visible duration. Automated discovery of API blueprints replaces manual imports across the codebase. Test-only API endpoints, such as POST /system/subconscious/tick, have been removed based on project rules. Compaction processing was unified, merging mid-ACT and end-turn logic into single processors. The pattern matcher was refactored from a large extractor into a single-pass LLM matcher. Memory seeding now occurs via a pre_act() hook, ensuring the audit row is recorded on every turn, even with zero recall.

• Per-tool start/end events + inline pill UI implemented for ACT loops

• Atomic dedup implemented in scheduler to prevent race conditions on scheduled_items

• PatternExtractor replaced by single-pass LLM matcher with scoped tools

• Memory pre-act seed now recorded via ToolRenderAndRecordService on every turn

• API blueprint discovery automated via pkgutil.iter_modules

• Plan-mode fire threshold lowered from 0.8 to 0.5 for productivity prompts