docs: operationalize complex-build authoring routes

[madawei2699]

Summary

• Add an Agent-facing semantic vs mesh/freeform vs hybrid decision record based on representation fit, not build size or landmark type.
• Document a reproducible OBJ → adapter-mesh → VoxelPlan workflow with source/artifact conventions, explicit scale/mode/budget/material options, hashes, bounded repair, fallback, and Luna-first escalation.
• Define bounded VoxelPlan hybrid composition and provenance requirements. Core currently lacks a checked composition helper; this PR documents the gap and does not add runtime composition machinery.
• Link the canonical guide from the authoring contract, Agent toolkit, llm.txt, and README.
• CraftDAG has no repo-local skill convention, so no parallel skill infrastructure was introduced. MinePilot Future: Evaluate Bedrock exporter feasibility for VoxelPlan targets #84 owns the local review bridge and should consume the route record, final VoxelPlan, and artifact/provenance manifest.

Validation

• pnpm build passed.
• pnpm test passed (218 tests).
• pnpm typecheck passed.
• pnpm lint passed.
• git diff --check passed.
• No runtime code, product primitives, benchmark artifacts, or MinePilot content were added.

Refs #123

[madawei2699]

Review of head acc0c8be1eeb136d242c0f581257c8f13c8e29c8: the docs-only workflow slice is well scoped and CI is green. I found no runtime/API blocker, but two points should be resolved before treating #123 as completed.

1. Do not require raw OBJ byte identity for every producer

COMPLEX_BUILD_AUTHORING.md currently says the same source/settings must regenerate the same OBJ bytes and treats OBJ byte differences as a reproducibility failure.

That is stronger than our own #114/#119 evidence supports: Blender exports for some statue/dragon cases were not byte-stable even when the canonical geometry / resulting VoxelPlan was deterministic.

Please narrow the contract:

• code-native/procedural exporters SHOULD produce byte-stable OBJ where practical;
• for producers whose serialization order is not stable, normalize/canonicalize geometry or record a canonical geometry hash;
• the hard production invariant is deterministic normalized geometry / VoxelPlan output for the pinned source+toolchain, not necessarily identical raw OBJ serialization bytes;
• raw OBJ SHA may still be recorded as provenance, but a changed raw SHA alone must not invalidate an otherwise canonical-equivalent deterministic result.

Avoid implementing a canonicalizer in this PR unless one already exists; this is primarily a documentation correction.

2. #124 completes the workflow-definition slice, not all #123 acceptance criteria

#123 also requires real complex-build evidence:

• one difficult large-building subsystem exercising mesh/freeform at practical Minecraft scale;
• one benchmark correctly remaining semantic;
• one hybrid attempt establishing whether composition is actually useful;
• Luna-class workflow validation.

Those are being handled by the parallel large-build benchmark/MinePilot work. Keep #123 open and keep this PR Draft until that evidence returns, or merge this as the canonical provisional workflow while explicitly leaving #123 open. Do not close #123 from this PR alone.

Everything else in this docs slice looks aligned with the intended boundary: representation-fit routing, Luna-first policy, no landmark-specific primitives, no Blender runtime dependency, and VoxelPlan as the common downstream target.

[madawei2699]

Final review of head 145977f4334a42018a9ba4f428d675ec35eb77a8: no remaining blocker.

The OBJ reproducibility contract now matches the actual #114/#119 evidence:

• byte-stable OBJ is preferred for code-native/procedural exporters;
• unstable producer serialization is allowed when canonical geometry remains equivalent;
• raw OBJ SHA is provenance, not the hard invariant;
• deterministic normalized VoxelPlan for pinned source/toolchain/options is the production invariant.

The PR remains correctly scoped to docs/workflow only. It does not add a premature composition helper, OBJ canonicalizer, Blender runtime dependency, or new semantic primitive. CI is green.

This PR is ready to leave Draft and merge. Keep #123 open until the merge lands; after merge, the current benchmark evidence supports closing #123 as completed with NO_HYBRID_JUSTIFIED recorded as the valid first complex-build routing result. MinePilot #84 remains the separate follow-up for unified local VoxelPlan review/capture workflow.