Discriminate Failed vs Empty Probe Cycles in run_cycle

[henry0816191]

Problem

The eval (Test 10) found that run_cycle conflates failed and empty probe cycles — a cycle that returns zero results because no new papers match the criteria is indistinguishable from a cycle that failed due to a network error, API timeout, or malformed response. Both produce the same return path. This is the core of the "Silent-Staleness Cascade" compound identified in the eval: when a probe cycle fails silently, the system falls back to stale data, and the operator cannot distinguish a genuine "no new papers" result from a failure that has stopped the pipeline. The health endpoint compounds this — it reports the last cycle's status but cannot tell the operator whether that status is genuine.

Acceptance Criteria

• run_cycle return type updated to a discriminated result (e.g., a dataclass/enum with Success(results=[...]), Empty(), and Failed(error=...) variants)
• Callers of run_cycle updated to handle all three variants explicitly
• Logging distinguishes between the three outcomes: successful cycle with N results, empty cycle (no matching papers), and failed cycle (with error detail)
• Metrics/counters (if any) track the three outcomes separately
• Health endpoint (if it reads cycle status) reports the discriminated result, not just "last run succeeded"
• Tests cover all three variants: successful, empty, and failed cycles
• Tests pass in CI (if applicable)
• PR approved by at least 1 reviewer

Implementation Notes

The simplest approach is a CycleResult dataclass with a status enum field (SUCCESS, EMPTY, FAILED) and optional results and error fields. The existing run_cycle function likely catches exceptions and returns an empty list — change this to return CycleResult(status=FAILED, error=str(e)) in the except block and CycleResult(status=EMPTY) when the probe returns zero results legitimately. This fix also addresses the "opaque failure path" planned item where config errors retry as transient — with a discriminated result, config errors can be classified as FAILED with a non-retriable error type.

References

• Eval finding: Test 10 (run_cycle conflates failed and empty cycles)
• Related files: run_cycle function (likely in src/scheduler.py or src/core.py), health endpoint handler, test files for cycle logic