feat: trajectory compression — learn from own best executions

[stackbilt-admin]

Summary

Port trajectory compression from the daemon (v1.65.0). Compresses successful LLM traces into few-shot examples per classification, injected back into dispatch prompts.

What to port

• Migration: 0022_trajectory_examples.sql
• Module: trajectory-compress.ts — trace grouping, quality scoring, compression
• Injection: getTrajectoryContext() called from augmentIntent() in dispatch.ts
• Registration in scheduled/index.ts

How it works

• Groups successful traces by classification (min 5 samples)
• Selects median-quality trace as representative example
• Quality score: confidence × completion length × inverse cost
• Injects as few-shot context: "Prior successful execution for similar query"
• 14-day TTL, max 20 examples, daily refresh at 04 UTC

Depends on

• LLM call tracing

[stackbilt-admin]

🤖 AEGIS: Queued as task ecd15cb8-aa20-4617-9e96-aadc6ac2615e (proposed). Awaiting operator approval before execution.

[stackbilt-admin]

🤖 AEGIS: Queued as task fc701805-7d36-4e93-9a97-0a10935ed565 (proposed). Awaiting operator approval before execution.

[stackbilt-admin]

Closing — the dependency model (aegis-oss#18) makes manual porting obsolete. Once internal AEGIS consumes @stackbilt/aegis-core as a dependency, this feature will flow downstream automatically because the core IS the production code.