test: MockProvider rejects transcripts a real provider would

[yuanhao]

The highest-leverage change available: it retroactively upgrades every existing test.

The structural hole

MockProvider::stream took _config: StreamConfig and discarded it. So 600+ tests drove the agent loop without a single one noticing when it built a message sequence every provider rejects.

That is not a coverage gap more tests would close. It is why loop detection reached a release branch injecting a steering message between an assistant's tool_use blocks and their tool_results — a history that shape poisons the agent, because it is kept, so every later prompt fails too. The suite was green throughout.

The change

The mock now checks the one invariant every provider enforces — a tool_use is answered by its tool_results before anything else intervenes — and panics with the reason a provider would give.

Panics rather than returning Err deliberately: a malformed transcript is a defect in the code under test, and a provider error would be swallowed by the very retry path that should never see it.

MockProvider::without_transcript_validation() exists for tests that build a malformed history on purpose. Nothing needed it.

Effect, measured

Reverting the 0.18 steer fix:

	before this PR	after
tests failing	0 — whole suite green	5, with the exact diagnostic

MockProvider received a transcript a real provider would reject:
a user message at [6] lands between an assistant's tool_use and its
tool_result — ["mock-tool-0"] still unanswered.

600+ existing tests became transcript guards without writing one.

One real fixture bug found

test_continue_from_tool_result resumed from [user, tool_result] — a tool result answering nothing. No provider accepts that, so it cannot arise in production, so the test proved nothing about the resume path it named. It now includes the assistant turn that made the call.

That it was the only failure across 629 tests is itself the useful signal: the loop's own transcript-building is clean, which is evidence the 0.18.1 fixes were right.

cargo test --all-features: 629 passed, 0 failed. Clippy clean under -Dwarnings.

Test-infrastructure only — no library behaviour change.

🤖 Generated with Claude Code