feat: lifecycle config + regression task model [LTV-Pn.3]

[shaypal5]

Summary

Third sub-PR of the split LTV-Pn (M6): the config + task-model plumbing the lifecycle pipeline needs, with no end-to-end wiring yet (that's Pn.4). Also discharges the long-deferred LTV-Pc regression-task-spec leftover.

GenerationConfig — lifecycle fields (layer: api)

New validated fields, consumed only by the lifecycle scheme (lead-scoring ignores them, just as the lifecycle side will ignore n_leads): n_customers, forward_windows_days (strictly-increasing tuple), early_tenure_weeks, observation_date (ISO string or None). Validation is isolated in _validate_lifecycle_fields. Kept flat on the shared config to match existing precedent (n_leads/snapshot_day); a nested per-scheme config is noted as a possible future refactor.

Lead-scoring data (tables/, tasks/) is byte-identical in both modes; only metadata/world_spec.json changes — by design, since it dumps the full config.

TaskManifest — regression support (layer: schema)

VALID_TASK_TYPES = {binary_classification, regression} with __post_init__ validation; docstrings broadened so label_column / label_window_days read target-agnostically.

Shared split writer (layer: render)

The deterministic shuffle/split/write logic is lifted to scheme-agnostic leadforge/render/tasks.py — target-agnostic (it never inspects the label), so it serves continuous pLTV targets. schemes.lead_scoring.render.tasks is now a thin wrapper defaulting the task — byte-identical tasks/ output. This repopulates the render.tasks path that LTV-Pf.2 vacated, now as shared envelope code (peer to render.manifests / render.relational_io); the module-layout lock test is updated to that intended end state.

Lifecycle task families (schemes/lifecycle/tasks.py)

lifecycle_task_manifests(regime) builds, per regime: three pltv_revenue_{90,365,730}d regression tasks + a churned_within_180d classification task. Calendar regime unprefixed; early-pLTV regime early_-prefixed so both families get distinct task directories. Targets/windows mirror the snapshot catalog so specs and columns can't drift. Completes LTV-Pc.

Tests (53 new; full suite 1835 passed / 51 skipped)

Config validation (bounds, sorted windows, ISO date); TaskManifest regression-type accept + invalid-type reject; shared writer on a continuous target (values preserved, deterministic, manifest emits regression); lifecycle task-family shape / cross-regime id uniqueness / prefixing / target-column match.

Next: Pn.4 — complete LifecycleScheme.build_world/write_bundle, lift the shared bundle orchestrator, first e2e lifecycle bundle.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds lifecycle-scheme configuration fields and regression-capable task manifests, plus a shared deterministic task-split writer that both lead-scoring and lifecycle can use (with lifecycle wiring deferred to the next sub-PR).

Changes:

• Extend GenerationConfig with validated lifecycle fields (n_customers, forward_windows_days, early_tenure_weeks, observation_date).
• Add task_type validation to TaskManifest and introduce VALID_TASK_TYPES including regression.
• Lift the deterministic shuffle/split/write logic into leadforge/render/tasks.py and make lead-scoring delegate; add lifecycle task-family definitions and targeted tests/docs updates.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/schemes/test_module_layout.py	Updates module-layout lock tests to reflect shared leadforge.render.tasks repopulation.
tests/schemes/lifecycle/test_tasks.py	Adds tests for regression task types and lifecycle task-family invariants.
tests/render/test_shared_task_writer.py	Adds tests for scheme-agnostic split writer, including continuous target preservation and determinism.
tests/core/test_config_lifecycle_fields.py	Adds coverage for new lifecycle config defaults and validation.
leadforge/schemes/lifecycle/tasks.py	Defines per-regime lifecycle task manifests (pLTV regression + churn classification).
leadforge/schemes/lead_scoring/render/tasks.py	Replaces lead-scoring split writer with a thin wrapper over the shared writer.
leadforge/schema/tasks.py	Introduces VALID_TASK_TYPES and validates TaskManifest.task_type; updates docs to be target-agnostic.
leadforge/render/tasks.py	New shared deterministic shuffle/split/write implementation for task exports.
leadforge/core/models.py	Adds lifecycle config fields and isolated validation in GenerationConfig.
docs/ltv/roadmap.md	Marks roadmap items as completed and documents the new lifecycle/task plumbing.
.agent-plan.md	Updates the running plan/status text for the LTV workstream.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

pr-agent-context report:

This run includes an unresolved review comment on PR #124 in repository https://github.com/leadforge-dev/leadforge

For each unresolved review comment, recommend one of: resolve as irrelevant, accept and implement
the recommended solution, open a separate issue and resolve as out-of-scope for this PR, accept and
implement a different solution, or resolve as already treated by the code.

After I reply with my decision per item, implement the accepted actions, resolve the corresponding
PR comments, and push all of these changes in a single commit.

# Copilot Comments

## COPILOT-1
Location: docs/ltv/roadmap.md:82
URL: https://github.com/leadforge-dev/leadforge/pull/124#discussion_r3407595317
Root author: copilot-pull-request-reviewer

Comment:
    The roadmap still documents the `task_type` enum as `regression | classification`, but the implemented/validated values are `binary_classification` and `regression` (see `VALID_TASK_TYPES`). Using `classification` here is misleading for contributors and readers.

Run metadata:

Tool ref: v4
Tool version: 4.0.21
Trigger: commit pushed
Workflow run: 27463528933 attempt 1
Comment timestamp: 2026-06-13T09:58:01.938335+00:00
PR head commit: 628ced51c20e9f204a7227d9c0f4b063c184c3e1