refactor: benchmark-driven general harness hardening

[dollspace-gay]

Scope

Establish honest Harness-Bench and Terminal-Bench baselines, then repair broad harness defects without task-specific shortcuts.

Current delivered fixes:

• honor documented global --dangerously-skip-permissions for ACP automation while preserving hard host safety boundaries
• accept standard ACP text content-block prompts while retaining legacy string compatibility and bounded validation

Integrity constraints

• upstream benchmark tasks, prompts, budgets, adapters, graders, and oracles remain unchanged
• no task-ID detection, expected-output inspection, benchmark-specific routing, or score manipulation
• fixes must improve the general harness and be validated outside the triggering benchmark task

Verification so far

• Rust 1.98 formatting and strict all-target/all-feature Clippy pass
• focused ACP, config, IDE, and CLI tests pass; known parallel fixture collisions remain tracked separately
• pinned acpx@0.13.2 real smoke reaches the provider through the documented global flag and standard ACP prompt shape

Current blocker

The configured Claude Agent SDK reports an expired OAuth session whose refresh failed. This is tracked as Crosslink #1238 and will be repaired without routing around the documented automation path.