fix(agent): add drop_session for cleanup paths; require base_sample in finish_session

[jingshenghang]

What

In the SWE coding-agent rollout, generate() calls adapter.finish_session(session_id) in its finally block as a catch-all cleanup. Since this path passes no base_sample, finish_session / get_trajectory defaulted base_sample=None.

On the normal path the trajectory is already consumed (the finally call is a no-op), but when the rollout raises before consuming the trajectory (e.g. sandbox boot, evaluate, or the wall-clock timeout guard fires), the session tree is still in the manager. The cleanup call then reaches to_sample(base_sample=None) and crashes on None.index.

Overloading base_sample=None to mean "cleanup only" also hid the intent: a reader can't tell from the call site whether it consumes the trajectory into trainable Samples or just frees state.

Changes

• TrajectoryManager.drop_session(sid) / BaseAdapter.drop_session(sid) — free a session's state (drain in-flight turns, pop the tree) without linearizing it into Samples. For paths that must release the session but produce nothing trainable.
• finish_session: base_sample is now required (was =None). The consume path can never silently build Samples from a None base; misuse fails loudly.
• generate.py finally: use drop_session(session_id) instead of finish_session(session_id) — the cleanup path now states its intent and can't crash.
• run_command (harness/common.py): the done marker holds the exit code as text, but echo N > done truncates the file before writing, so a poll landing in that window reads "" and int("") raised. Skip empty reads and keep polling instead.

Behavior

• Success path: unchanged — finish_session(base_sample=...) consumes and returns Samples; the finally drop_session is a no-op (tree already popped).
• Abort/timeout path: drop_session releases the session cleanly, no crash, nothing trained.