feat(agent): streamline minimal harness prompt

[nonoqing]

Summary

Streamline the stable Minimal Harness system prompt while preserving the cross-tool coding and safety contracts that are not expressible through individual tool schemas.

• Start with the concise persona You are a helpful software engineer assistant.
• Remove duplicated per-tool documentation from the system prompt; tool parameters and single-tool behavior remain owned by the request-time descriptions and schemas.
• Retain repository-instruction discovery, read-before-edit, focused changes, truthful verification, untrusted-content handling, destructive-action safeguards, and defensive-security boundaries.
• Keep unavailable Goal, Task, Subagent, Web, MCP, dedicated Grep/Glob, and deferred-gateway capabilities out of both the prompt and the Minimal manifest.
• Update the architecture design and prompt catalog contract test to lock the concise policy.

Type and Areas

Type:

Feature, prompt-policy refinement, test, and architecture documentation.

Areas:

Agent content assembly, Minimal Harness prompt policy, prompt catalog contracts, and Minimal Harness architecture documentation.

Motivation / Impact

The previous Minimal prompt repeated detailed tool instructions already supplied by the live tool descriptions and schemas. This increased prompt size and constrained model behavior without adding runtime enforcement.

The revised prompt keeps only stable persona and cross-tool rules that the schema or runtime cannot communicate on their own. Minimal remains a BitFun product profile with the same tool manifest, permissions, execution lifecycle, provider/model configuration, and stable prompt key. Balanced and other profiles are unchanged.

Verification

• cargo test --locked -p bitfun-agent-content --test prompt_catalog_contracts — 4 passed, 0 failed.
• Headless Linux DeepSWE/Pier sample campaign deepswe-bitfun-dsv4flash-max-minimal-prompt10-20260818-bitfun-0f850c7-pier-47adce3 — all 10 trials completed with 0 agent exceptions; official verifier reward was 7/10. This validates live Minimal startup and execution for this sample, not broad product behavior or a final benchmark score.

Reviewer Notes

• This PR contains one commit and changes only three files.
• No model ID, provider, reasoning preset, sampling configuration, permission policy, tool manifest, runtime limit, or adapter behavior changes are included.
• The stable policy key remains minimal-harness-v1; only its prompt bytes and matching contract/documentation change.
• The server exercise covered the local/headless CLI path inside Pier isolation. It is not evidence for Remote Workspace, Remote Control, Peer Device Mode, or Detached Dispatch behavior.
• AI-assisted: yes. Testing level: focused contract test plus a 10-case live-model evaluation sample; broad cross-surface suites were not run.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.