feat: add event-boundary timing instrumentation for PoW

[vitrixLab]

Summary

This draft PR implements the Phase 4 PoW timing instrumentation for AIOpsLab without changing the existing TTD, TTL, TTA, or TTM result semantics yet.

What changed

• Adds EvaluationTiming with explicit lifecycle events:
  • fault_occurred
  • detection_completed
  • localization_completed
  • analysis_completed
  • mitigation_completed
• Records the fault boundary immediately after fault injection.
• Records detection/localization/RCA completion when a valid submission is accepted.
• Defers mitigation completion until the task evaluator confirms successful recovery rather than treating submit() alone as proof of mitigation.
• Persists timing events in session JSON/W&B output.
• Adds a focused timing unit-test module covering explicit elapsed intervals and missing-boundary behavior.
• Adds an evaluation-timing design note documenting the intended event-boundary formulas and the historical-compatibility rule.

Why

The current session duration is a useful legacy measurement but does not encode the lifecycle boundaries needed for research-aligned timing metrics. This PR establishes the event provenance first so a later metric-migration PR can safely reclassify timing fields after all task types and no-op cases are covered by regression tests.

Compatibility

Existing metric fields are intentionally preserved. Historical session-duration results are not rewritten or relabeled by this PR.

Validation

• Branch is one commit ahead of main and has no merge/rebase/reset effects.
• Unit tests were added for the new timing primitive.
• Full repository test execution was not run in this environment because there is no local AIOpsLab checkout and the GitHub CLI is unavailable here.

Follow-up

The next PoW step should add task-level integration tests for the event boundaries, especially mitigation-oracle completion and no-op problems, before changing the semantics of the existing metric fields.

[vitrixLab]

GE / RAV — PoW Evidence-Boundary Test

Downgrading this from a proposed PR-level governance artifact to a controlled PR comment test.

RAV scope: comment/evidence behavior only.

• Claims are treated as claims, not validation evidence.
• Deployment/preview evidence does not by itself establish functional acceptance.
• No security or governance conclusion is inferred beyond the evidence actually observed.
• No merge authorization is implied.
• No repository mutation is requested or performed by this test.

Result: PASS — INFORMATIONAL / EVIDENCE-BOUNDARY TEST

CLAIM ≠ VALIDATION ≠ ACCEPTANCE ≠ AUTHORIZATION

This PoW demonstrates the governance/evidence-handling process independently; it does not represent upstream microsoft/AIOpsLab approval or acceptance.