Add null-byte delimited stdin prompt mode to Omni/Agent

Goal

Make Omni/Agent accept prompts from stdin when no script argument is given. Delimiter is a single null byte (\0) between prompts. EOF on stdin = clean exit.

Design

Current flow: agent script.md parses frontmatter + markdown, pushes initial user message, runs agentLoop once, exits.

New flow:

• If an argv script is provided → current behavior unchanged (shebang stays happy).
• If no argv script → initialize empty session state (or from frontmatter on stdin first if we want), then enter a read-prompt → run-loop → read-prompt cycle:

1. Read bytes from stdin until a \0 is encountered. 2. Decode as UTF-8, push as a user message into session state. 3. Call agentLoop with the same state; it runs until no more tool calls / natural termination. 4. When the loop returns, reset asHitMaxIterations / other per-turn flags, go to step 1. 5. EOF on stdin → exit 0.

Acceptance criteria

• echo -ne "hello\0" | agent runs one loop iteration and exits.
• printf "first prompt\0second prompt\0" | agent runs two loops, preserving session state between them (context accumulates, conversation continues).
• agent script.md still works exactly as before (shebang case is untouched).
• All existing Trace.Event emission works the same; agentd's replay/debug tools can consume the output unchanged.
• Session context persists across prompts (this is the whole point — agent stays hot, cache warm).
• Compaction still runs between iterations via existing Compaction.compactMessages hook.

Files likely to touch

• Omni/Agent.hs — main entry point and agentLoop
• Possibly Omni/Agent/Session.hs or wherever session state init lives

Notes

• Don't add a --mode or --interactive flag. Stdin-reading mode is activated by absence of a script argument. This is how python, ghci, bc, sh all work.
• Don't add JSON command parsing. A prompt is literally <utf8-bytes>\0. Nothing else.
• Later subtasks handle signals and the daemon swap.