fix(orchestrator): use active session's provider/model when injecting prompts

[ryuhaneul]

Background

_inject_prompt in orchestrator/injection.py builds the AgentRequest
without provider_override or model_override, so CLIService falls
back to the bot's default provider/model.

Sessions are per-topic (SessionKey(transport, chat_id, topic_id)), so a
single chat can have one topic on Codex and another on Claude even when
the bot default is claude. When the default disagrees with the topic's
active session provider, --resume <sid> hits the wrong CLI:

CLI stderr: No conversation found with session ID: <sid>
CLI returned empty output (exit=1)

The empty result becomes the envelope's result_text and surfaces in
the user chat as the visible reply, even though the task completed
normally (status=done or failed). The same path covers async
inter-agent result injection, so sub-agent results into a
different-provider topic hit it too.

Same family as #82/#81 (d38d738), which added a stale-session marker
and retry to handle_interagent_message. The _inject_prompt path was
not covered.

Proposed change

Reuse the already-resolved active session and pass its provider/model:

provider_override=active.provider if active else None,
model_override=active.model if active else None,

When active is None, both stay None and behavior is unchanged.

Consistency

Other resume sites already pass per-session provider/model — e.g.
flows.named_session_flow uses provider_override=ns.provider, model_override=ns.model. The inject path was the odd one out.

Test plan

• Added tests/orchestrator/test_injection.py mocking get_active
  and asserting the captured AgentRequest uses the active
  session's provider/model (and falls back to None when no active
  session).
• ruff check, ruff format --check, mypy clean.
• Manual repro on a Codex-active topic with Claude default,
  status=done task. Before: No conversation found … in user
  chat. After: result delivered cleanly.