Implement persistent native agent state

[safranchik]

Summary

This change makes each mn.Parameter the persistent native state of one owner agent.

• Forward creates a disposable episode fork from a read-only Parameter.
• Backward resumes the episode and accumulates structured directional feed.
• DFM.step() resumes each persistent owner once with all accumulated feed.
• Owner updates run in parallel and promote as one atomic model generation.
• Shared, recurrent, retained, and frozen Parameter graphs follow PyTorch-shaped behavior.
• Pi, Codex, Claude Code, OpenCode, Hermes, and Prime Agent use isolated native state.
• The FFT discovery example provides an exact, resumable research workload.

Lifecycle

forward()                 temporary episode forks
loss.backward()           accumulate Parameter.feed
optimizer.step()          one owner reduction per Parameter
optimizer.zero_feed()     clear accumulated feed

Forward and backward cannot mutate canonical Parameter state. Episode state is never merged or promoted. Git remains private to HyTorch and provides atomic model generations.

Safety

• Keep credentials outside persisted Parameters.
• Reject nested Git metadata, escaping symlinks, and special files.
• Limit native state to 100,000 files and 2 GiB.
• Preserve feed after a failed reducer so the caller can retry or clear it.
• Roll back the complete step candidate if any owner reducer fails.

Validation

The complete offline suite passed on Python 3.11, 3.12, and 3.13:

• 111 passed
• 2 opt-in real-provider tests skipped
• uv lock --check
• uv run ruff check .
• uv run ruff format --check .
• uv build
• git diff --check