test: fix stale/environment-fragile tests (nightly check-and-test)

[yilu331]

Companion to ReflexioAI/reflexio-enterprise#240, opened by the nightly check-and-test run.

These tests passed in isolation but failed in the full run because they were stale relative to intentional behavior changes, or fragile to the CI environment:

• test_env_loader_mode: isolate ambient BACKEND_PORT — the loader is override=False, so a BACKEND_PORT exported by the suite's shell wins over the file value.
• test_unified_search_floor: enable the relevance floor explicitly; its default was flipped to off in fix(server): harden runtime defaults #161.
• test_embedding_service_concurrency: assert serialized model inference (peak == 1); fix(embedding): serialize model inference to stop concurrent-encode tensor race #153 added _MODEL_ENCODE_LOCK to stop the concurrent-encode tensor race, so encodes no longer overlap.
• test_recycle_smoke_integration: raise the /healthz readiness timeouts — a recycled worker reloads the embedder + cross-encoder, which exceeds the old budget on a contended box.
• test_interaction_workflows: add a timeout(300) marker so cold embedder + reranker loads under the default -n auto run don't spuriously hit the 120s per-test timeout.

Also includes ruff-format normalization of two source files.

Verified: the fixed tests pass with and without the ambient env that originally broke them.

Summary by CodeRabbit

• Tests

  • Enhanced test robustness by isolating environment variables and strengthening assertions for concurrency control and unified search configuration.
  • Increased timeout budgets for end-to-end workflows and integration tests to accommodate model loading under resource constraints.

• Refactor

  • Restructured internal billing constant and reformatted code for improved maintainability.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 01a0a3e9-c10a-46a0-9140-af52354552a9

📥 Commits

Reviewing files that changed from the base of the PR and between 424649c and 7235e73.

📒 Files selected for processing (7)

• reflexio/server/billing_meter.py
• reflexio/server/services/extraction/outcome.py
• tests/cli/test_env_loader_mode.py
• tests/e2e_tests/test_interaction_workflows.py
• tests/server/llm/test_embedding_service_concurrency.py
• tests/server/services/test_unified_search_floor.py
• tests/server/test_recycle_smoke_integration.py

---

📝 Walkthrough

Walkthrough

The _INTERNAL billing caller constant is changed from a string to a tuple. ExtractionOutcome.completed is cosmetically reformatted. Three tests are hardened for correctness: env isolation before loading a mode env file, explicit enabled=True in a floor config, and a stricter peak-encode-concurrency assertion of exactly 1. Four timeout budgets are extended to accommodate slow model loading in CI.

Changes

Billing, extraction, and test fixes

Layer / File(s)	Summary
Production: billing constant and extraction outcome formatting reflexio/server/billing_meter.py, reflexio/server/services/extraction/outcome.py	_INTERNAL is changed from a string literal to a single-element tuple while its usage as caller_type is unchanged. ExtractionOutcome.completed return is reformatted to multi-line with no behavioral change.
Test correctness: env isolation, floor config, embedding lock assertion tests/cli/test_env_loader_mode.py, tests/server/services/test_unified_search_floor.py, tests/server/llm/test_embedding_service_concurrency.py	test_mode_filename explicitly removes BACKEND_PORT from os.environ before the env-load call. test_floor_applied_per_arm sets enabled=True explicitly in RetrievalFloorConfig. test_embed_texts_caps_and_queues tightens the peak-concurrency assertion to exactly 1 to match model encode lock serialization semantics.
CI timeout increases for model-loading tests tests/e2e_tests/test_interaction_workflows.py, tests/server/test_recycle_smoke_integration.py	E2e interaction workflow pytestmark adds pytest.mark.timeout(300) for cold model-load scenarios. The recycle smoke test increases initial startup /healthz polling from 45 s to 90 s and post-recycle polling from 15 s to 90 s.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

• ReflexioAI/reflexio#153: Introduced the shared model encode lock that serializes embedding inference, directly motivating this PR's change of the peak-concurrency assertion to exactly 1.
• ReflexioAI/reflexio#126: Introduced _INTERNAL = "internal" in billing_meter.py; this PR changes that constant to a tuple.
• ReflexioAI/reflexio#60: Introduced the daemon-mode recycling smoke test and /healthz-based readiness polling whose timeout values are extended in this PR.

Poem

🐇 A tuple instead of a string, what a curious thing,
The tests now know to clear the port before they sing,
Peak concurrency locked to one — fair and square,
Ninety seconds for models to load, with room to spare,
Small fixes stitched together with a rabbit's care! 🎉

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch ci/check-and-test-nightly-fixes

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}