Separate live Senpai skills from operator guides

[morganmcg1]

Summary

Stacked on #3498. This PR isolates the Senpai runtime-skill boundary from the program.md system-prompt work.

• make plugins/senpai/skills the sole source for Senpai-owned live skills: assignment, delegation, result submission, review, human Issues, status, Exa, AlphaXiv, and W&B
• keep onboarding, operator, legacy workflow, and developer guides under .agents/skills; move bootstrap-target there
• stop copying runner .agents/skills into pods and disable ambient user-skill loading for root and file agents
• resolve file-agent skill declarations only from the filtered target-project catalog, never from user skill directories
• attach the Senpai plugin directly to every root and process-isolated child OpenHands conversation
• continue loading explicit target-repository skills through the separate project-owned channel while excluding repository instruction files
• update runtime helper paths, entrypoint checks, docs, and integration tests for the new ownership boundary

This prevents interactive onboarding and developer-only skills from reaching live autoresearch agents while keeping every live Senpai skill in the plugin OpenHands loads. The separate grilling-autoresearch skill is in sibling PR #3501; Claude skill discovery is isolated in sibling PR #3502.

Verification

• focused runtime skill/plugin/Exa suite — 140 passed
• full test suite after refreshing from Add program.md to system context and clean up docs/ #3498 — 931 passed, 6 skipped
• bash -n on the modified entrypoint/context scripts
• git diff --check
• git diff --exit-code origin/main -- paper

Merge order

Merge #3498 first, then retarget or rebase this PR onto main before merging it. PRs #3501 and #3502 are independent sibling stacks on #3498.