test(predict-rlm): consolidate core regression suite

[eriberdy022]

Rationale

The default package suite had accumulated overlapping backend scenarios, repeated inherited tests, static/default/export checks, presentation assertions, and mock-only plumbing. This refactor concentrates coverage on observable regression contracts. CI also exposed a JSPI deadline bug: interrupts could abort asyncio scheduler callbacks and orphan execution results instead of returning recoverable timeouts. The runtime fixes preserve public APIs while correcting that failure.

Summary

• Reduce collected cases from 1,475 to 389 (74%) and Python test lines from 37,418 to 12,206 (about two thirds).
• Consolidate shared backend contracts and remove duplicate inherited SBX cases, obsolete fixtures, static/presentation checks, and one-off benchmark/example checks from the default suite.
• Preserve RLM execution, typed predictions, cancellation/resource ownership, file/workspace safety, trace/accounting, Codex auth/transport/cache, and GEPA acceptance/evidence/resume regressions.
• Synchronize SBX callbacks across worker event loops with a thread-safe barrier. Spawn isolated timeout-reproduction processes instead of forking multithreaded pytest; separate startup from execution budgets.
• Scope JSPI deadline tracing to generated code; cancel suspended execution tasks without raising through scheduler callbacks; acknowledge watchdog disarm before cleanup; restore the previous SIGINT handler.
• Add a red-before-green awaited-sleep timeout regression with state and signal-handler preservation. Keep prediction preparation outside deliberately short loop deadlines and disable unrelated eager package installation in lightweight adapter/deadline tests.
• Preserve CI selections while adding unbuffered per-case JSPI logs and stack dumps. Update architecture, changelog, and test-policy documentation. Example-local suites remain separate.

Test Plan

• Latest HEAD a96ac237254edb711fa6fc8aedde8db0058c96bf: workflow succeeded, with all eight jobs green: core on Python 3.11/3.12/3.13, SBX, Codex LM, GEPA, JSPI integration, and lint.
• Reproduce the SBX cross-loop failure with PYTHONASYNCIODEBUG=1; verify synchronous and asynchronous overlap after correction.
• Demonstrate the new awaited-sleep regression failing before the scheduler fix and passing afterward.
• Verify final synchronous-sleep, RLM state/history/predict, and asynchronous-sleep deadline scenarios under Python 3.12, fresh dependencies, Deno 2.9.6, and coverage: 3 passed.
• Initial full-suite refactor validation, before CI follow-up fixes: uv run --all-extras pytest -o addopts= -v --tb=short --junitxml=/tmp/predict-rlm-refactor-final.xml — 358 passed, 30 skipped.
• Ruff passes for src/ and tests/; changed Python tests are formatted.
• Run opt-in real SBX and bootstrap Docker scenarios in provisioned environments. These prerequisites were not enabled locally; other skips reflect unsupported backend capabilities.

The initial local full run reported DSPy deprecations and async-stream cleanup warnings. No warning suppression, retries, or test deselection were added to obtain a green workflow.