feat(v1): add Prime Agent ACP harness

[hallerite]

Summary

• add Prime Agent as a normal v1 ACPHarness with MCP, resume, and skills support
• point Prime Agent's Chat Completions provider at the Verifiers interception endpoint
• make autonomous continuation opt-in and forward the rollout turn, token, and timeout budgets as Prime Agent CLI arguments
• cover Prime Agent in the shared ACP resume/MCP end-to-end matrix
• build directly on feat: support RLM training over ACP #2386 so nano-RLM and Prime Agent share the same generic ACP contract

There is no Prime-specific lifecycle protocol or runner in Verifiers. The standard
ACP session/prompt response is the authoritative completion boundary; Prime Agent
is responsible for returning it only after its own causal work is quiescent.

Companion Prime Agent fix

Depends on PrimeIntellect-ai/prime-agent#1612, which:

• waits for terminal causal quiescence before completing an ACP prompt
• propagates descendant failures and the final stop reason through that request
• preserves opaque Chat Completions reasoning details during same-model replay

Merge gate: after the Prime Agent fix lands and is published, update the pinned
PrimeAgentHarnessConfig.version. The current beta predates that fix; the
exact-source run below validates the intended combined stack.

Validation

• Seth's exact feature canary against Prime Agent source d940d75 and this Verifiers
  head:
  • all skill, continual-memory, child lifecycle, explicit messaging, retained-child
    follow-up, deletion, answer, and cleanup checks passed
  • the ACP request stayed open until the child was deleted
  • the canary's aggregate audit is false only because it requires the intentionally
    omitted private trace.info.acp_lifecycle telemetry
• live shared ACP resume/MCP E2E passed against the companion source stack
• autonomous CLI forwarding verified for disabled/default and enabled budgeted modes
• focused trace, idempotency, and RLM contract tests pass
• pre-push markdownlint, Ruff, format, and Ty checks pass