Add program.md to system context and clean up docs/

[morganmcg1]

Summary

program.md system context

• add the optional program_path launch setting for advisor and student roles
• when blank, search the repository root and one directory below it, proceeding only when exactly one program.md exists
• fail before constructing a controller or model worker when discovery has no match or is ambiguous, listing every match and explaining how to set program_path
• append the selected file to every main and child system prompt under # program.md - <path>
• keep the resolved path and formatted section together as ProgramSystemPrompt, nested inside one immutable SenpaiSystemInstructions value with the loaded harness, selected role, and authoritative launch context
• load all four instruction sources once when each agent process starts; remove system-context digests, refresh prompts, rereads, and reinjection, and use GitHub Issues for live human direction
• retain only a small StartedConversationLedger so the initial controller context is not replayed after a successful first turn

Launch configuration

• rename the ambiguous runner settings to senpai_repo_url and senpai_repo_revision, with matching SENPAI_REPO_* pod variables and no legacy aliases
• keep the public Senpai source as a read-only default while making clear that target_repo_url is the required write and PR destination
• combine runtime and isolation rules in one rendered SENPAI-LAUNCH-CONTEXT.md and append it after program.md in the system instructions
• transport optional operator instructions separately as initial user context, never as trusted launch policy
• fail loudly on missing, malformed, or empty launch context and explicitly propagate the exact resolved context to delegated children

Prompting and onboarding

• remove target prompt-advisor.md / prompt-student.md loading, templating, bootstrap guidance, and compatibility; program.md is the sole repository-specific research contract
• define Python-assembled model instructions directly as importable string constants in senpai_agent/PROMPTS.py, retain strict one-pass interpolation, and remove the Markdown parser and runtime prompt-file loading
• keep the checked-in harness and Advisor/Student role charters as source documents while composing their loaded text, program.md, and authoritative launch context through the single aggregate
• keep program.md guidance in role prompts and live skills, derive metrics and research context from it rather than physical-AI assumptions, and keep the shared harness free of program.md references
• add the live delegate-subagents skill for agent/model selection and the spawn/await/status/cancel workflow; remove duplicated mechanics from Advisor/Student prompts and keep hard tree limits in the shared harness
• align Advisor assignment guidance with the typed create_assignment contract by putting the complete experiment brief in body
• document the student-only post_assignment_comment question/progress channel conditionally until its implementation in draft PR Restore student PR comments #3494 lands; keep terminal publication on submit_experiment_result
• stop loading repository AGENTS.md, AGENT.md, or CLAUDE.md instruction files as OpenHands project context; explicit target skills remain available through a filtered skill path
• trim the shared harness to Senpai-specific control-plane rules by removing project-instruction discovery, generic skill-catalog, launch-context/time, and completion reminders
• remove controller-only “target” vocabulary, move event-handling guidance into the shared Senpai harness, and keep active prompt prose unwrapped
• add concise program.md authoring guidance and four examples to AGENTS.md, exposed to human Claude users through the root CLAUDE.md symlink; add a five-section Mission/Data/Evaluation/Files/Research outline to the README
• remove the retired docs/ source tree and deployment workflow while retaining the README link to the published project site; paper/ stays unchanged because every needed figure already exists there

Three independent follow-ups stack directly on this PR: #3500 separates live runtime skills from human/developer guides, #3501 adds the grilling-autoresearch onboarding skill, and #3502 exposes project skills to Claude.

Verification

• quick_validate.py plugins/senpai/skills/delegate-subagents — valid
• focused student-prompt, harness, and agent-boundary suite — 65 passed
• focused prompt, package, program, OpenHands, controller, delegation, and launch suite — 226 passed
• full test suite — 938 passed, 6 skipped
• built a clean wheel and verified it contains senpai_agent/PROMPTS.py without the removed Markdown catalog or lowercase loader module
• git diff --check
• git diff --exit-code origin/main -- paper