Make C12 qualification deterministic

[KeyffMS]

Summary

Replaces the model-planned C12 setup with a deterministic runtime probe so project_doc_max_bytes is tested independently of planner behavior.

C12 automatic-loading probe

• creates a 4,320-byte AGENTS.md with a secret head marker inside the first 1,024 bytes and a secret tail marker at byte 3,072;
• sets project_doc_max_bytes = 1024 in project config and redundantly passes -c project_doc_max_bytes=1024 to codex exec;
• requires a tool-free read-only response: command-tool use blocks the probe;
• the prompt contains marker names but never marker values, so the tail value cannot be guessed from the prompt;
• outer harness marks C12 FAILED if the secret tail marker beyond the budget is actually visible.

PlanAnvil explicit-read assertion

• runs the real map_instructions.py implementation outside the model;
• verifies the full 4,320-byte size, complete SHA-256, full_read=true, and truncation_risk=true at the same 1,024-byte automatic budget;
• computes the final C12 decision deterministically without an evaluator-agent.

Scope / safety

• C12 is the only newly overridden capability;
• existing C01–C11 and C13–C16 deterministic fixes are inherited unchanged;
• no sandbox bypass or danger-full-access;
• full qualification remains workflow_dispatch-only from main;
• this PR uses only standard hosted CI.