Add canonical US runtime stage contracts, manifests, and docs

[anth-volk]

Summary

Implement the canonical 9-stage US Microplex runtime taxonomy proposed in #20.

This issue tracks adding first-class stage contracts, saved-run manifests/overlays, resume-oriented stage artifacts, and docs for the canonical US build path.

Scope

• Define the canonical 9 runtime stages in code.
• Add stage contracts that describe each stage's inputs, outputs, artifacts, diagnostics, validation placeholders, and resume posture.
• Write a saved-run stage_manifest.json alongside artifact bundles.
• Add stage-local artifacts where the current pipeline already has stable boundaries, including source planning, PE entity materialization, calibration summary, dataset assembly, and validation evidence.
• Update data-flow snapshots and benchmark evidence backfills to use the canonical stage IDs.
• Add docs for the pipeline stages, stage contracts/manifests, and API reference using the repo's Sphinx/Jupyter Book docs path.

Out of scope for this first implementation

• Automatic conditional execution or a scheduler.
• A shared stage-validation execution engine.
• Reworking the underlying modeling/calibration method.
• Full source-frame snapshotting for every raw input source.

Acceptance criteria

• Saved artifact bundles expose the 9 canonical stages.
• Stage manifests are machine-readable and stable enough for saved-run overlays.
• Existing artifact save and checkpoint evidence paths remain backward compatible.
• Stage documentation is available in the repo docs.
• Focused tests cover stage contract ordering, manifest shape, data-flow snapshot integration, and checkpoint evidence integration.

Related discussion: #20