fix(pi): keep runtime context at system boundary

[Tarun-joy]

Summary

• inject Pi runtime routing, active memory, and resume context through before_agent_start.systemPrompt
• keep automation at Pi's host-owned model-only boundary
• remove the custom/user message injection path so runtime instructions cannot appear as user-authored chat

Why

Pi's display: false only controls the interactive transcript. API consumers and adapters may still reinterpret custom messages as user messages. Runtime automation is system-owned context, so it belongs in the system prompt rather than in conversation history.

This intentionally favors transcript correctness and stable role semantics over the previous prompt-cache optimization.

Validation

• npx vitest run tests/pi-extension.test.ts — 68 passed
• focused Resume/Routing tests — 7 passed
• tests assert no context handler is registered
• clean scoped diff

[Tarun-joy]

Review verdict: APPROVE

The revised implementation uses Pi's canonical before_agent_start system-prompt result instead of relying on display: false for role correctness. This keeps context model-visible, prevents API/adapters from reclassifying automation as user chat, preserves the existing prompt prefix, and removes pending-message state plus the one-shot context hook.

Validated with the complete Pi adapter suite (68/68), focused Resume/Routing coverage (7/7), typecheck/build and bundle drift checks. The branch is clean, scoped to the Pi adapter and tests, current with upstream main, and mergeable. No blockers found.

[Tarun-joy]

Revalidated after rebasing onto current upstream main (da14dc1).

• Pi adapter suite: 68/68 passed
• Typecheck: passed
• Build + bundle integrity + asymmetric drift: passed
• Branch: clean and MERGEABLE / CLEAN
• Overlap rechecked: Hide Pi context injection from user entry #864 still uses hidden custom messages, which Pi/API adapters can reclassify as user-role content; this PR keeps runtime automation at the canonical system boundary instead.