fix(acp): keep team prompt before agent persona

[johnely19]

Summary

• frame shared team instructions before the per-agent system prompt
• keep agent-specific core, huddle, canvas, and turn content after the persona
• extend the exact-byte framing regression control to pin the cache-shaped order

Evidence

• negative control: the focused framing test failed when the implementation was temporarily restored to System → Team while the expectation remained Team → System
• env -u BUZZ_ACP_LAZY_POOL -u BUZZ_ACP_IDLE_POOL_SLEEP cargo test -p buzz-acp — 830 library tests + 9 lifecycle tests passed on tree 8a73a5471b1c247f887b779a527a7124910589f7

Live measurement status

The source trace and diff are complete. Before/after real GLM-turn measurements remain blocked: accepted real turns reproducibly fail with the adapter's authentication reply before provider usage is returned. Direct calls using the same inherited credential, endpoint, mapped models, and Bearer-header shape return HTTP 200, so credential rotation is explicitly not requested; the remaining failure is isolated to the Claude Agent SDK/Claude Code execution path. The patch is not installed in the Desktop-owned active harness, and no live cache improvement is claimed from the local test.

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is eed74bde2f4797714335ac10c56c0b0244c1def4...bd95611426c43983bb153c47b46e01642ff5c32d.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review bd95611426c43983bb153c47b46e01642ff5c32d to authorize a new review.
Any previous review applies only to its recorded range.