test(harness): shared test fakes + conformance determinism fix

[declan-scale]

Summary

First slice of #425 (post-merge harness cleanup), scoped to non-breaking test infrastructure so it can land independently.

• Shared test fakes — extract tests/lib/core/harness/_fakes.py (FakeSpan/FakeTracing), removing ~9 duplicated in-file copies.
• Conformance determinism fix (Greptile P1) — test_span_derivation_is_deterministic now iterates all_fixtures() at runtime, paired with a new conformance conftest.py that eagerly imports every per-harness conformance module so the fixture set is fully populated regardless of collection order. Previously import-time parametrization silently dropped per-harness coverage.
• Removes the now-landed pr4-pydantic-ai planning doc.

Test plan

• pytest tests/lib/core/harness/ — 122 passed, 1 skipped

Notes

Part of a stacked split of #425. No source changes; safe to merge first.

🤖 Generated with Claude Code

Greptile Summary

This PR is a pure test-infrastructure cleanup with no source code changes. It extracts shared FakeSpan/FakeTracing test doubles into tests/lib/core/harness/_fakes.py, eliminating roughly nine in-file duplicate definitions. It also fixes a conformance determinism gap: test_span_derivation_is_deterministic previously parametrized on all_fixtures() at import time (missing any harness whose module hadn't been collected yet) and now calls all_fixtures() inside the test body, backed by a new conftest.py that eagerly imports every per-harness conformance module so registration is always complete.

• Shared fakes — FakeSpan / FakeTracing extracted to _fakes.py; all local _FakeTracing / _RecordingTracing / _FakeSpan variants replaced with the shared versions throughout test_tracer.py, test_emitter.py, test_auto_send.py, test_yield_delivery.py, and runner.py.
• Determinism fix — test_conformance.py::test_span_derivation_is_deterministic is now a single no-parametrize test that calls all_fixtures() at run time, with a guard asserting that per-harness fixtures were registered; conftest.py ensures all conformance modules are imported before any test executes.
• asyncio.run() at import removed — test_pydantic_ai_conformance.py and test_codex_conformance.py now use the shared run_pure_async loop-free driver; the driver itself is promoted from a private copy in test_claude_code_conformance.py to a public export in runner.py.

Confidence Score: 5/5

Safe to merge — no source code changes, only test infrastructure refactoring with a determinism bug fix.

All changes are confined to test files. The shared FakeTracing is a behavioral superset of every local copy it replaces (same started/ended tuple format, same span-object shape). The run_pure_async driver is sound for pure in-memory coroutines and was already proven correct by the existing claude_code_conformance suite. The conftest eager-import strategy is safe because Python's module cache prevents double-registration. No edge cases that could silently hide failures were found.

No files require special attention.

Important Files Changed

Filename	Overview
tests/lib/core/harness/_fakes.py	New shared test-doubles file; FakeSpan/FakeTracing match the superset of all previously-local variants and expose started_names/started_pairs/ended_outputs convenience properties for backward-compatible assertions.
tests/lib/core/harness/conformance/conftest.py	New conftest eagerly imports all five per-harness conformance modules for registration side-effects; test_openai_conformance.py confirmed to exist; no asyncio.run() in any imported module.
tests/lib/core/harness/conformance/runner.py	Removes local _FakeTracing (replaced by shared FakeTracing), promotes run_pure_async from test_claude_code_conformance.py to a public export, and cleans up stale ticket-reference comments.
tests/lib/core/harness/conformance/test_conformance.py	test_span_derivation_is_deterministic converted from parametrize-at-import to a runtime loop with a guard ensuring per-harness fixtures are registered; sound fix for the collection-order determinism gap.
tests/lib/core/harness/conformance/test_pydantic_ai_conformance.py	asyncio.run() replaced with run_pure_async for all four module-level fixture builds; per-harness determinism test removed (consolidated into test_conformance.py).
tests/lib/core/harness/conformance/test_codex_conformance.py	asyncio.run() in _build() replaced with run_pure_async; local span-derivation test removed and covered by the consolidated test in test_conformance.py.
tests/lib/core/harness/test_tracer.py	_FakeSpan/_FakeTracing removed; FakeTracing imported from _fakes.py; shared implementation is an exact behavioral match to the removed local version.
tests/lib/core/harness/test_auto_send.py	Local _RecordTracing removed; assertions updated from .started/.ended to .started_names/.ended_outputs to match the richer shared FakeTracing tuple storage.
tests/lib/core/harness/test_yield_delivery.py	Local _RecordTracing removed; assertions updated to started_names/ended_outputs; otherwise identical behavior.

_{Reviews (5): Last reviewed commit: "test(harness): build conformance fixture..." | Re-trigger Greptile}

[declan-scale]

@greptile review

[declan-scale]

@greptile review