test: make #57's trigger fire for the person who trips it

[yuanhao]

Closes #57 by moving its conditional somewhere it will actually be noticed.

Why not just leave the issue open

#57 proposed replacing Content::ToolCall.provider_metadata's untyped Option<serde_json::Value> with a typed, namespaced enum. Its own "when to do this" was a conditional:

Trigger: the moment a second provider needs provider_metadata — do both items together before shipping that provider, since that is when the key-collision risk becomes real.

An open issue is a poor mechanism for a conditional. Whoever adds OpenAI reasoning IDs will not check the tracker first, so the reminder fires for nobody — which is what happened: open since 2026-07-05, doing nothing.

This test fires for exactly that person, at exactly that moment, in the diff that trips it:

A provider outside the Gemini family now writes `provider_metadata`:

  openai_responses.rs:210 — provider_metadata: Some(serde_json::json!({"reasoning_id": "r-1"})),

... Do #57 first, then add the field.

The issue's analysis is preserved in the test's module docs, including why now is the last cheap moment: after a second family ships, persisted Session JSONL carries the untyped shape and the migration inherits a serde back-compat burden it doesn't need.

What it actually checks

Deliberately not a grep for provider_metadata: Some. The Gemini production sites assign a variable (provider_metadata: metadata), and the Some literals in those files are test fixtures — so matching on Some would guard test code and miss the real writes. The invariant that holds today is the stronger one: every provider outside the Gemini family writes the literal None.

Two exclusions, both earned by the guard failing on them first rather than assumed up front:

• A struct field declaration is not a write — mock.rs declares the field on its fixture type.
• mock.rs echoes whatever MockToolCall was handed, so it forwards metadata rather than originating any. It cannot be the second family this watches for.

Verification

Real trigger simulated	openai_responses.rs writing a reasoning_id → fails, naming file, line and remedy
Guard reads nothing	asserted against — a layout change that found <5 provider modules fails rather than passing green
Second Gemini key	also a trigger, since that is where namespacing starts to matter

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

Test-only — no library change.

🤖 Generated with Claude Code