Fix #615: OTel tracer leak from test_full_pipeline pollutes test_ping in full-suite runs

[Axiya3749]

Fixes test isolation: an earlier test's leaked global state made 4 test_ping.py tests fail only when the full suite runs, training readers to ignore a red suite.

Summary

• Restore tokenjam.sdk.agent._tracer to its original value in full_stack's teardown instead of leaving it bound to a shut-down provider and closed DB.

Root cause

full_stack (tests/integration/test_full_pipeline.py) rebinds the module-global agent_mod._tracer to a tracer on its own local TracerProvider, then calls provider.shutdown() and db.close() in teardown — but never restores _tracer. Because tests/integration collects before tests/unit alphabetically, this leaks: test_ping.py's record_llm_call() calls then run through the dead TjSpanExporter/closed DuckDB (raising "Connection already closed", logged as "Span export failed" at provider.py:56) instead of the real global provider that test_ping's own proof exporter is attached to. The proof exporter never captures anything, so intercepted is wrongly False and the 4 tests asserting on "intercepted a test span" / interception-dependent output fail.

Fix mirrors the existing correct pattern in tests/agents/test_record_outcome.py: save the original tracer before rebinding, restore it in teardown, before the provider is shut down.

Related issue

Closes #615

Tests / Verification

• pytest tests/unit/test_ping.py -v — 10/10 pass
• pytest tests/integration/test_full_pipeline.py -v — 11/11 pass
• pytest tests/ — full suite: the 4 previously-failing test_ping tests are fixed; confirmed via git-stash isolation that a pre-existing, unrelated failure in tests/unit/test_deadweight.py (a Rich line-wrap issue in _render_deadweight, unrelated to OTel/tracer state) is present with or without this change, both standalone and in the full suite

What's NOT in this PR

• No change to tokenjam/ source — this is purely a test-isolation fix in tests/integration/test_full_pipeline.py.
• tests/unit/test_deadweight.py::test_render_deadweight_names_the_dead_server remains failing — confirmed pre-existing and unrelated (fails standalone, independent of this fix and of test ordering). Separate root cause in the CLI renderer; not touched here.

Checklist

• Tests pass (pytest tests/unit/ tests/synthetic/ tests/agents/ tests/integration/)
• Lint clean (ruff check tokenjam/)
• Type check clean (mypy tokenjam/)
• CLAUDE.md updated (if architecture changed) — N/A, no architecture change
• Test spans use tests/factories.py (not raw NormalizedSpan) — N/A, no spans added
• Requested @anilmurty as reviewer (or @-mentioned him above)

@anilmurty

[greptile-apps]

Greptile Summary

Restores the SDK module’s original tracer during full_stack fixture teardown before shutting down the fixture-local provider and database.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
tests/integration/test_full_pipeline.py	Saves the pre-test tracer and restores it during teardown, preventing fixture-local OTel state from leaking into later tests.

_{Reviews (3): Last reviewed commit: "Merge branch 'main' into fix-issue-615" | Re-trigger Greptile}

[anilmurty]

Thanks @Axiya3749 — verified the root cause and your fix: the full_stack fixture rebinds the module-global agent_mod._tracer and never restores it, so a later test_ping routes through the dead exporter / closed DuckDB. Reproduced 4 failures on main → 0 with this patch. Clean, correct, mirrors the save/restore pattern already in test_record_outcome.py. Merging.