Add preview_correction_impact workflow proposal with rule-based rationale

[akgohain]

Motivation

• Provide a backend-only proposal that estimates retraining readiness from recent proofreading/correction activity.
• Make recommendations transparent and rule-based by using explicit constants for thresholds and an export window.

Description

• Add server_api/workflow/proposals.py implementing PROPOSAL_TYPE_PREVIEW_CORRECTION_IMPACT and build_preview_correction_impact_proposal(...) which summarizes correction event counts, recent exports, and returns a recommendation plus a human-readable rationale that references explicit constants (LOW_CORRECTION_THRESHOLD, HIGH_CORRECTION_THRESHOLD, and RECENT_EXPORT_WINDOW).
• Introduce CORRECTION_EVENT_TYPES and EXPORT_EVENT_TYPES and a deterministic recent_exports summarizer that does not mutate input state.
• Export the new API from server_api/workflow/__init__.py for clean imports.
• Add tests/test_workflow_correction_preview.py covering low-correction and high-correction scenarios and asserting payload shape, recommendation behavior, and that inputs are not mutated.

Testing

• Ran PYTHONPATH=. uv run --no-project pytest -q tests/test_workflow_correction_preview.py which passed with 2 passed.
• Note: an initial uv run pytest -q tests/test_workflow_correction_preview.py failed in this environment due to local editable dependency metadata for pytorch_connectomics, which does not affect the unit tests when PYTHONPATH=. is used.
• Tests confirm the proposal includes summary + recommendation + rationale and that there is no direct state mutation.

---

Codex Task

[akgohain]

Consolidated into draft PR #165 (feat/codex-wave-integration). Closing this source draft to keep review in one place.