Add Claude observability parity and prepare v0.2.0

[tmusser]

Summary

Brings the Claude harness up to the same observability contract as the merged Codex checkpoint path and prepares the repository for a v0.2.0 release.

Claude parity

• Capture distinct Claude workspace states after completed file-changing tool results.
• Briefly pause the Claude process group while copying stable snapshots.
• Defer public and hidden evaluator execution until after Claude exits.
• Keep hidden evaluator output outside the agent session.
• Select the initial evaluator for initial runs and the resume evaluator for full/stripped resume runs.
• Record complete/partial coverage, stable snapshot status, skipped states, snapshot pause overhead, evaluator overhead, and native observation unit.
• Keep polling fallback conservative: stable/deferred snapshots are retained, but first-green remains non-conclusive because sampling can miss states.
• Refuse exact stream coverage when only a final snapshot exists or any file-changing tool result is unresolved.

Provider capability rubric

Adds docs/capabilities.md with a Claude/Codex matrix covering:

• native observation units;
• stable snapshots;
• deferred evaluation;
• first functional and bench-ready green;
• post-green tails;
• coverage guarantees and fallback boundaries;
• cross-provider comparison guidance.

The README includes a short provider-observability callout and link.

v0.2.0 preparation

• Bump package version from 0.1.0 to 0.2.0.
• Update the README release badge.
• Add v0.2.0 changelog notes and explicit claim boundaries.

The tag/release should be created from main after this PR merges; this PR deliberately does not create a tag from an unmerged branch.

Adversarial boundary

Parity means equivalent evidence standards, not interchangeable units. Claude turns/tool-result boundaries and Codex provider items remain distinct. Claims about the first evaluator-green observed boundary require checkpoint_coverage_complete=true; neither path claims an instruction-level instant. Post-functional work is not automatically waste when it completes required bench-ready evidence.

Validation

• Focused Claude/Codex parity tests: passed.
• Fake-Claude red-to-green end-to-end test: passed.
• Final-only and capped-coverage negative tests: passed.
• Full benchmark_harness/tests suite: passed.
• Shell syntax and git diff --check: passed.
• Release metadata checks: passed.
• Final repository CI: passed on Python 3.11 and Python 3.12.

The final diff contains only the 11 intended implementation, test, documentation, and release-metadata files.

[github-actions]

Claude parity focused tests failed:

F.........................                                               [100%]
=================================== FAILURES ===================================
_ test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage _
benchmark_harness/tests/test_claude_solution_latency_parity.py:85: in test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage
    assert summary["first_functional_green_turn"] == 2
E   assert 1 == 2
=========================== short test summary info ============================
FAILED benchmark_harness/tests/test_claude_solution_latency_parity.py::test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage - assert 1 == 2
1 failed, 25 passed in 3.47s

[github-actions]

Claude parity focused tests failed:

F.........................                                               [100%]
=================================== FAILURES ===================================
_ test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage _
benchmark_harness/tests/test_claude_solution_latency_parity.py:85: in test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage
    assert summary["first_functional_green_turn"] == 2
E   assert 1 == 2
=========================== short test summary info ============================
FAILED benchmark_harness/tests/test_claude_solution_latency_parity.py::test_stream_json_observer_reports_first_green_turn_with_complete_stable_coverage - assert 1 == 2
1 failed, 25 passed in 3.58s

[github-actions]

Claude parity hygiene validation failed:

== bash syntax ==
bash_exit=0
== git diff --check ==
docs/scorecard.md:118: new blank line at EOF.
diff_exit=2

[tmusser]

Final validation update: the earlier red→green attribution failures and EOF hygiene failure were intermediate staging findings and are superseded. The cleaned head ccb35a045aa0fc2d8bf3b86f76e476dd6074dafb passes standard CI on Python 3.11 and 3.12. Focused Claude/Codex parity tests, the full harness suite, shell syntax, diff hygiene, release metadata, and negative coverage tests are green. Final diff: 11 intended files only.