First step of the peer-schemes architecture (LTV-M2). Introduces the
generation-scheme abstraction and routes Generator.generate() through it,
wrapping the existing lead-scoring pipeline as the first registered scheme.
Lead-scoring output is unchanged (the wrapper delegates to the identical
functions); this PR is a pure refactor.

- leadforge/schemes/base.py — GenerationScheme protocol, SCHEME_REGISTRY,
  register_scheme/get_scheme/available_schemes, UnknownSchemeError. The
  protocol covers the generation half (build_population + simulate) wired
  through Generator; render dispatch is added later (LTV-M6).
- leadforge/schemes/lead_scoring/__init__.py — LeadScoringScheme, delegating
  to leadforge.simulation.{population,engine} (relocated under this package in
  LTV-Pe); self-registers on import.
- leadforge/schemes/__init__.py — imports built-in schemes for their
  registration side effect; re-exports the registry API.
- Recipe gains a `scheme` field (default "lead_scoring"); parsed in from_dict.
  b2b_saas_procurement_v1/recipe.yaml declares `scheme: lead_scoring`.
- WorldSpec gains `scheme` (default "lead_scoring"); from_recipe threads
  recipe.scheme through; Generator.generate() resolves and runs the scheme via
  get_scheme() instead of calling build_population/simulate_world directly.

World-graph sampling and bundle rendering still run in their current paths
(lead-scoring only registered); they move under the scheme abstraction in
later M2/M6 PRs.

Tests: tests/schemes/test_registry.py (15) — registry resolution, protocol
conformance, conflict/unknown handling, Recipe.scheme parsing/validation,
WorldSpec default, Generator threading + end-to-end dispatch. Full suite 1495
passed / 51 skipped; ruff + mypy clean (87 files).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Check off LTV-Pd, link PR #107, note the byte-identical verification, and
advance the agent-plan status to LTV-Pe next.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…V-Pd]

Acts on a hostile self-review of the first revision. The original cut the
abstraction at build_population/simulate with lead-scoring-shaped signatures
(world_graph, category_latent_correlations, latent_touch_intensity,
PopulationResult/SimulationResult), leaving graph sampling, difficulty
interpretation, and bundle assembly hardcoded in Generator.generate(). That
seam could not accommodate a non-lead-scoring scheme — defeating the point of
extracting it early.

Changes:
- GenerationScheme protocol is now a single `build_world(config, narrative,
  **options) -> WorldBundle`. No lead-scoring types leak into the contract;
  scheme-specific flags ride through **options.
- LeadScoringScheme.build_world owns the whole lead-scoring pipeline: hidden-DAG
  sampling, difficulty-profile → DifficultyParams + category-latent
  correlations (extracted to _resolve_difficulty), population, simulation, and
  WorldBundle assembly. Generator.generate() is now scheme-agnostic
  (override resolution + dispatch only).
- Fix dropped-scheme bug: build_world sets spec.scheme=self.name, so
  bundle.spec.scheme reflects the actual scheme (was always the default).
- DEFAULT_SCHEME constant in core.models removes the duplicated "lead_scoring"
  literal across Recipe/WorldSpec; a test guards LeadScoringScheme.name ==
  DEFAULT_SCHEME.
- Tests: regression for bundle.spec.scheme, determinism through the scheme
  path, name/default drift guard; protocol-conformance test annotated as the
  weak structural check it is.

Verified still byte-identical to main (14/14 files of a pinned-timestamp
bundle). Full suite 1498 passed / 51 skipped; ruff + mypy clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>