fix(agent): wait for session file in agent send --wait

[codeaholicguy]

Closes #400

Repro

ai-devkit agent start --type pi --name test-agent --cwd /tmp/somewhere
ai-devkit agent send --id test-agent --wait --timeout 120000 --json "hello"
# -> ✖ No session file found for agent "test-agent"

Root cause

The agent registry entry is created at agent start, but the session file only materializes when the harness writes its first entry — which happens on the first message. sendToAgent resolved agent.sessionFilePath at dispatch time and threw No session file found immediately when it was absent, so send --wait right after start could never succeed.

Fix design

In sendToAgent (packages/cli/src/services/agent/agent.service.ts):

• The prompt is sent first (unchanged delivery path via sendAgentPrompt). This matters: the send is often what triggers the harness to write its first session entry.
• When agent.sessionFilePath is absent, the new waitForSessionFile polls manager.listAgents() every 500ms until the file appears, bounded by min(--timeout, 30s) default grace.
• If the agent disappears from listAgents() while polling, a distinct error is raised: Agent "<name>" exited before creating a session file — separate from the generic response timeout.
• If the grace elapses: Agent "<name>" did not create a session file within <N>ms.
• Session-file appearance doubles as the harness readiness signal, covering the related "send into a still-booting TUI is silently dropped" symptom noted in the issue.
• When sessionFilePath already exists, behavior is unchanged: the transcript baseline is still seeded before the send so no messages are skipped.

Validation

• Unit: 3 new/updated cases in agent.test.ts — file appears after send → success; agent exits → distinct error; grace expiry → bounded error. packages/cli suite: 1341 tests pass.
• e2e: new e2e/agent-send-wait.e2e.ts drives the compiled sendToAgent through a scripted listAgents() sequence reproducing start-then-send-wait. Trade-off: a real agent start boot needs tmux + a live harness binary, too heavy for CI — so process detection and tmux delivery are stubbed via the injected runtime/focusManager seams while the send → poll → wait path runs against the real built dist. Verified it catches the bug: all 3 tests fail against the pre-fix code.
• npm run lint, npm test, npm run test:e2e all green (pre-commit hook + manual runs).

Generated with Devin