review fixes: integration.yml gates + bare-except + env-var + drift-tolerant skip

[saurabhjain1592]

Summary

Deep code review (cross-SDK audit prompted by user's question "what about Python and Go?") caught several issues in the Python SDK that the original QF-10 pass excluded because Python was tagged "already ported, don't touch." Same audit lens that produced TS PRs #196/#197/#199 + Java PR #147 + the just-merged TS+Java+Go review-fixes PRs.

P0 — same shape as TS+Java+Go logs ordering

1. integration.yml logs-after-teardown ordering. `Stop community stack` (`if: always`) declared BEFORE `Show docker logs on failure` (`if: failure`). On a failed run, `compose down --volumes --remove-orphans` destroyed the containers; subsequent `compose logs` returned empty. Fixed: capture logs first, persist to disk, upload as artifact, then teardown.

P1 — workflow + correctness

2. integration.yml no concurrency group. Back-to-back pushes spawn parallel docker-compose stacks fighting over ports 8080/8081. Fixed: added concurrency: integration-${{ github.ref }} with cancel-in-progress: true.

3. integration.yml cron Monday → Tuesday 06:00 UTC. Aligns with TS + Java + Go integration crons; failures land in EU/IN working hours.

4. integration.yml no actions/upload-artifact for logs. Even with the ordering fixed, when the runner is gone the logs are gone. Fixed: persist docker compose logs --tail=500 to disk before teardown, upload as artifact on failure.

5. integration.yml docker compose up -d --wait --wait-timeout 120 for native dependency-aware health gating. Kept the existing curl /health loop as belt-and-suspenders.

6. integration.yml no pip cache. All three setup-python@v5 steps re-installed deps from cold every run. ci.yml already set cache: 'pip'; integration.yml didn't. Fixed: matched ci.yml's cache config across all three jobs.

7. examples/openai_integration.py two bare except Exception: blocks. Both printed (expected if no OpenAI key) for every error class, masking SDK regressions, schema drift, and governance failures. Fixed: narrowed to openai.OpenAIError / PolicyViolationError. Anything else now bubbles up.

8. examples/wcp_retry_idempotency.py env-var name disagreement. Used AXONFLOW_BASE_URL; the rest of the SDK and every other example use AXONFLOW_AGENT_URL. The script worked only because its own fallback default kicked in. Fixed: renamed to AXONFLOW_AGENT_URL.

9. tests/test_integration.py test_generate_plan over-broad skip markers. Used string-matching on (\"LLM\", \"provider\", \"Planning Engine\", \"not available\", \"not initialized\") to decide whether to skip — silently absorbed any error containing those words, including unrelated SDK regressions. Fixed: gated on explicit AXONFLOW_HAS_PLANNING=1 env var instead of pattern-matching error text.

What I checked and did NOT flag

• Decommissioned URLs: Python is genuinely clean. No staging-eu, no fictional hostnames, no v1.17.0 module-path traps. Sandbox helper correctly defaults to localhost:8080.
• pyproject.toml BLE001 in examples/**: flagged P2 by the review agent — would catch future bare excepts in examples. Deferred for a separate small PR; doesn't affect current correctness.
• AXONFLOW_TRY=1 mode test coverage: flagged P1 by review agent — needs a new test file, not a one-line fix. Tracking separately.

Test plan

• Contract tests: 21/21 pass (`pytest tests/test_contract.py`)
• All touched Python files compile cleanly (`python -m py_compile`)
• CI green on PR
• CI: integration green on push to main (basic + gateway examples + new ordering + new artifact upload)

CHANGELOG

User-facing entries under `[Unreleased] / Fixed` for openai_integration narrowing, env-var rename, and test_integration drift-tolerant skip.