Remove the live-provider test tier

[anassg-lago]

Not a response to a review comment — a decision to delete a test tier that never ran anywhere.

Neither CI workflow sets a single provider key, so tests/integration always skipped. Running it needed five paid accounts. And each per-provider test pointed at an in-process mock Lago, so it could not prove Lago accepts an event even when it did run. None of it ever shipped to clients either — the package is built from src/ only.

Adapter behaviour stays pinned by the captured real responses under tests/unit/adapters/fixtures/, which is what the unit tests assert against; CONTRIBUTING now says that explicitly so the removal doesn't read as "provider shapes are unverified". Live verification moves to a driver script pointed at a real Lago — the only way to exercise what these tests couldn't: llm_cost against a dynamic charge, and precise_total_amount_cents being accepted.

Reviewing this one is just checking that nothing else referenced the directory: git grep tests/integration comes back empty outside CHANGELOG history.

[ancorcruz]

Review — dropping the live tier

Agreed on the substance: eight of the nine files are skipif-gated on provider credentials, CI runs only tests/unit, so they were dead weight in the repo and the captured fixtures are the better contract. No objection to those going.

One file in the set isn't a live-provider test, flagged inline: test_outage_replay.py has no skipif and needs no credentials — it spins a local HTTPServer on 127.0.0.1 that toggles 503→200. It runs anywhere, and it covers exactly the queue paths #18 rewrites.

Separately, and worth fixing at the repo level: none of these three PRs has run CI. ci.yml triggers on pull_request: branches: [main], which filters on the PR's base. All three PRs in this stack target other branches in the stack, so there are zero workflow runs on any of them:

feature/cloudflare-gateway-connector   runs=14
chore/drop-live-test-tier              runs=0
fix/queue-subscription-and-pricing     runs=0
fix/gateway-and-gemini-attribution     runs=0

That means the four blocker fixes in #18 have no automated verification at all. Splitting the work into a stack was the right move, but it silently opted the whole stack out of CI — the one thing that made #13's size tolerable.

I ran the CI job locally against each branch in turn. All three are clean:

branch	ruff	format	mypy	pytest
chore/drop-live-test-tier	pass	pass	pass	513 passed
fix/queue-subscription-and-pricing	pass	pass	pass	531 passed
fix/gateway-and-gemini-attribution	pass	pass	pass	533 passed

Dropping the branches: [main] filter from the pull_request trigger (keeping it on push) would cover any stacked PR from here on.

[ancorcruz]

Keep this one — it's the only file in the set that isn't a live-provider test, and it covers what #18 changes.

Checked all nine: eight carry a skipif gated on provider or Lago credentials. This one has no skipif and no credentials — _spawn() starts a local HTTPServer on 127.0.0.1:0 whose failing flag returns 503, so both tests run offline, in-process, in about a second.

What they assert is the queue's contract under an outage: 200 events buffered through a failure window arrive in order and complete (test_outage_replay_preserves_order_and_count), and a long outage at the buffer cap drops the oldest then drains the rest (test_long_outage_at_buffer_cap_drops_oldest_then_drains).

That is the same surface #18 reworks — _PERMANENT_STATUSES, _send_individually, _replay_failed ordering, the overflow path. #18's own unit tests are good and cover the specific regressions, but they drive EventQueue directly with fake senders; this is the only thing in the repo that exercises buffer → HTTP → retry → recovery end to end against a real socket. Deleting it in the PR immediately beneath the one that rewrites those paths is the wrong order.

Since CI only runs tests/unit, this file has never actually run in CI either — so the fix is to move it rather than keep it where it is. tests/unit/test_queue_outage.py (or a tests/local/ that CI also runs) gets it executing for the first time. It takes ~1s and needs no network.

The other eight: no argument, they should go.

[anassg-lago]

You are right, and I got this one wrong. I lumped all nine files together as "the live tier" without checking each for a skipif — test_outage_replay.py has none, needs no credentials, spins a local HTTPServer and runs in about a second. Deleting it in the PR directly beneath the one rewriting _PERMANENT_STATUSES, _send_individually and _replay_failed ordering is exactly the wrong order, and it was the only thing in the repo exercising buffer → HTTP → retry → recovery against a real socket.

Restoring it as tests/unit/test_queue_outage.py, which also gets it running in CI for the first time. Two notes while doing that: the same file exists in the JS stack and its deletion is still unmerged, so I will drop that deletion from the open JS PR rather than delete-then-restore there; and one of its two tests asserts oldest-dropped-first at the buffer cap, which is now also covered by the FIFO-ordering test #18 added — worth keeping both, since only this one goes through a socket.

[anassg-lago]

On CI: you are right, and it is the more serious of the two. ci.yml filtering pull_request on branches: [main] meant the whole stack ran zero workflows — so the four blocker fixes in #18 merged with no automated verification. Splitting for reviewability silently removed the thing that made the split safe. Thank you for running the matrix locally.

Dropping the branches filter from the pull_request trigger (keeping it on push) in the same follow-up as the outage-test restore, so any stacked PR is covered from here on. I will also re-run the full suite against the merged feature/cloudflare-gateway-connector tip, since none of the three PRs was ever validated in its merged form — only branch by branch.