test: repair P1 legacy suite baseline

[Pigbibi]

Summary

• make the existing legacy pytest gate fail on real regressions instead of suppressing its exit code
• align stale heartbeat, profile rollout, switch-plan, locked-uv, and scheduler assertions with current source-of-truth behavior
• isolate the two consecutive_losses tests from shared temporary performance state
• mark the local test package so third-party tests packages cannot shadow it

Root cause

• 9 failures were stale assertions left behind by already-merged notification localization, QSL runtime-enabled profile metadata, locked-uv workflow, and scheduler continuity changes
• 2 failures depended on shared temp-state records; a clean runner correctly returned no consecutive-loss history
• the test directory lacked __init__.py, allowing eventkit's top-level tests package to win import resolution

No production strategy, risk, broker, order, provider, replay, deployment, scheduler execution, QPK, or dependency pin behavior is changed.

Validation

• uv sync --frozen --extra test
• existing legacy suite with frozen pins and isolated TMPDIR: 417 passed, 5 skipped
• same legacy suite with the exact CI external refs (QPK 748b623, UES 7c2f036, UESP 2dd46de): 417 passed, 5 skipped
• focused 11-failure regression set plus CI contract: 12 passed
• isolated consecutive_losses regression: 2 passed
• uv run --no-sync ruff check --exclude external .
• uv lock --check
• uv run --no-sync python scripts/check_qpk_pin_consistency.py
• git diff --check
• git diff --exit-code -- pyproject.toml uv.lock qsl.toml
• actionlint -color (exit 0; reports only two pre-existing shellcheck warnings in untouched workflows)

Scope

This PR is independent from #454. It does not modify that PR's branch or commits and does not merge or delete any branch.