[Bug]: [Presets] Any failure during resume marks the creation permanently failed

[peterschmidt85]

Steps to reproduce

1. Start a creation from a configuration with any env entry, e.g. env: [MODEL_LABEL=some-value].
2. Stop it: dstack preset stop <id> -y. It correctly records interrupted and prints Resume it with dstack preset resume <id>.
3. Resume from a shell where that variable is not set: dstack preset resume <id>. It fails with Environment variable MODEL_LABEL is not set ([Bug]: [Presets] Resume requires literal env values in the local environment #4165).
4. Resume again, with the variable set: dstack preset resume <id>.

Actual behaviour

Step 4 refuses:

Preset <id> creation failed and cannot be resumed

Step 3 is what marked the session failed. create_preset wraps the whole creation in except BaseException: _close_agent_session(session, "failed") (create.py:437-439), so any exception is terminal: it writes failed and deletes the workspace (create.py:754-756, workspace.py:120-130). Step 3's exception is raised at create.py:415, before _create_preset is entered, so nothing had started and nothing had changed, yet the session became unresumable. load_resumable_session then refuses (session.py:331-332).

Not caused by reconcile: _is_reconcilable requires status running, an existing report, and a dead owner (create.py:252-262), so an interrupted session is never reconciled.

Expected behaviour

Step 3 leaves the session interrupted, and step 4 resumes it. A failure that happens before the agent is spawned is not a failed creation.