Add prompt compression to all agent prompts

[msoedov]

Agent prompts (system prompt + API ref + rules + issue context) are assembled in internal/scheduler/scheduler.go and sent verbatim on every run. There is no compression or deduplication across runs.

Problem:
Boilerplate (rules, API ref, wiki guidelines) is repeated identically in every invocation. For agents running on short cron cycles, this is wasteful.

Questions to resolve:

• Which prompt sections are static vs dynamic?
• Can static sections be cached via prompt caching (provider-dependent)?
• Is there a generic compression approach that works across all runners (claude, gemini, codex, copilot, opencode)?
• Should this be a build-time optimization (minify prompts) or runtime (cache/hash)?

Acceptance criteria:

• Identify and quantify static vs dynamic prompt portions
• Implement compression for at least one runner
• Measure token/cost reduction