docs: clarify inbound task reply routing

[zzz163519]

Summary

• document how an inbound task's current-turn final is routed back to its caller
• warn agents not to open a second ask merely to report completion
• clarify --chain, --silence, correction, and provider-vs-business completion semantics
• project the same rules into managed provider memory and every bundled ask skill

Why

An agent can finish practical work, return to its prompt, and separately message the original caller without terminalizing the inbound turn. That leaves the original job and mailbox head active. Clear runtime instructions reduce this failure mode while the daemon-side orphan diagnostics in #260 are designed separately.

The user guide also explains that managed AGENTS.md/CLAUDE.md content is materialized for provider sessions and should not be assumed to hot-reload into an existing session.

Related: #260, #261

Validation

• /home/calvin/tmp/claude_codex_bridge/.venv/bin/python -m pytest -q test/test_ask_skill_templates.py test/test_project_memory.py
  • 17 passed
• /home/calvin/tmp/claude_codex_bridge/.venv/bin/python -m pytest -q test/test_provider_memory_external_matrix.py test/test_provider_memory_external_context.py test/test_project_memory_real_context.py
  • 1 passed, 2 skipped
• git diff --check

PDF generation was not run because latexmk is not installed in the test environment.

[zzz163519]

CI note: the full Tests matrix is currently red on upstream main as well. The latest main run (29558401022) and this PR's Ubuntu 3.10 job both fail in the unchanged lifecycle test test_ccbd_stop_all_does_not_run_post_shutdown_heartbeat (destroy_events == []), after the rest of the suite passes. This failure is outside this PR's diff; the PR-specific targeted/smoke checks are reported in the description.