test(integration): allow skill reuse across turns

[moonyue-w]

Summary

Allow the Forward and Managed resource integration scenarios to reuse skill content from an earlier turn. The second turn still verifies the generated skill marker and a successful idle state, while the first turn still requires a tool event and the correct file and environment values.

Why

A model can read the mounted skill during the first turn and return its marker in the second turn without another tool call. Requiring a new tool event in both turns incorrectly fails that valid execution.

Verification

• uv run make lint
• uv run make typecheck
• uv run make test
• uv run make docs-check — no API documentation changes; covered by PR CI.
• Example --help smoke checks — examples unchanged.
• Account-backed integration tests run, or not required — CN/Global validation will follow the merge and internal synchronization.

Commands and results:

uv run --extra dev --locked pytest tests/test_scenarios_offline.py -q
28 passed
uv run make lint
Passed
uv run make typecheck
Passed (605 source files)
uv run make test
3114 passed, 21 account-backed integration tests deselected

The six added offline cases cover both modes: a correct second-turn marker without a new tool event passes, missing first-turn tool execution fails, and an incorrect second-turn marker fails. Before the scenario fix, both valid cached-skill cases failed with "No actual tool execution".

Impact

• Public API or behavior changed
• User-facing changes added to CHANGELOG.md under Unreleased, or not applicable — test-only change.
• Synchronous and asynchronous clients remain aligned — SDK implementations unchanged.
• Generated API documentation updated — not applicable.
• Contract fixtures updated — not applicable.
• Go and TypeScript parity considered — this fixes the Python resource scenario's per-turn assertion.
• Migration notes added for a breaking change — not applicable.
• No credentials, .env.live, logs, or generated test output committed

Additional context: This PR contains only the two Python resource scenarios and their offline regression tests. Internal CI Dream exclusions and version preparation are separate changes.