test(server): refactor spec_compat_hooks tests to use public API

[bokelley]

Follow-up to #646 / PR #648.

The tests in tests/test_spec_compat_hooks.py (added in PR #648) reach into private helpers (_hook_get_products, _infer_asset_type, _coerce_asset, _make_sync_creatives_hook) and assert on internal shapes. Per CLAUDE.md the testing contract is:

• Test what you import — call the public API, not internals.
• Validate with .model_validate() — prove hook output is actually 4.4-schema-valid by feeding it through the generated SyncCreativesRequest / GetProductsRequest models from adcp.types.

Scope

1. Rewrite assertions to go through spec_compat_hooks() + the published hook dict surface (e.g. via create_tool_caller with the hooks installed), not the private helpers.
2. Add at least one test that runs the coerced sync_creatives payload through SyncCreativesRequest.model_validate() and asserts no ValidationError.
3. Expand the create_tool_caller integration test (currently only checks the buying_mode default on get_products) to exercise the harder paths:
   • format_id wrap (string → object)
   • asset_type inference for unknown assets
   • image → url demotion through the actual sync_creatives tool path
4. Optional: emit a warnings.warn when exclude= contains a name not in the canonical set, so typos like exclude={"sync_creative"} don't silently leave the hook active.

Why not block PR #648 on this

The implementation in src/adcp/server/spec_compat.py is correct and matches the spec table from #646. The current tests do exercise every code path (just at the wrong level); regression coverage is real. Splitting the test refactor out keeps the feature shippable.