Land build-manifest contract + release-manifest emitter into src/microplex_us (deferred until MP settles)

[MaxGhenis]

Land the build-manifest legibility contract + release-manifest emitter into src/microplex_us

Reviewed, tested work from the scoreboard-first pass on branch
claude/scoreboard-first-rebuild (under rebuild/). Deferred until Microplex
settles/launches; land coordinated with Codex's live main.

Two pieces, both absent from main today, both independently reviewed (review-fix
cycle clean) + tested:

• Build-manifest contract — rebuild/docs/build-manifest-spec.md +
  rebuild/manifest/build_manifest.schema.json + worked example. Per-dataset
  provenance/legibility: spine-vs-donor split, per-variable treatment
  (kept / imputed_on_clone / overridden_everywhere / imputed_from_donor / derived),
  environment capture, reproduce command + output binding. The governance layer
  main lacks (and the answer to "what exactly is this build, and how do I rebuild it").
• Release-manifest emitter — rebuild/mp_rebuild/release_manifest.py. Emits a
  policyengine-bundles DataReleaseManifest (validated against the live bundles
  JSON schema and pydantic model, certified + uncertified). This is the
  bundle-eligibility artifact mp lacks today (eCPS has utils/release_manifest.py).
  Fed by the build manifest so provenance composes into certification.

Land in src/microplex_us/ (e.g. release_manifest.py), wired into the build
pipeline + the existing mp gates / benchmark manifest. Not the
comparison/scoreboard from that branch — that is validated-redundant with main's
ecps_replacement_comparison (two independent implementations converged on the
same sound design, which was the point of the parallel run).

Tests on the branch: 31 fast + 7 manifest-conformance passing.

[MaxGhenis]

Updated per Data-weekly (2026-05-29)

The build-manifest is now the agreed live architecture contract, extended to match the meeting:

• calibration_diagnostics.targets[] — per-target aggregate outputs (target_value / estimate / relative_error / source / in_loss_function / geography / pipeline). Closes the "Microplex stores no per-variable aggregate outputs anywhere" gap raised by Pavel; the columns are the calibration-dashboard columns. Emitted incrementally → a live dashboard can render build progress.
• run — unique non-trampled run_id, emitted_at_stage (live emission), and an upload policy: default auto-upload to a central store (Supabase, PolicyEngine org), opt-out incognito for purely-experimental local runs (so results stop getting lost).

This is the contract the calibration dashboard + the query API/CLI build against, with the stage-emit writing the incremental snapshots — i.e. it composes with, doesn't compete with, that work.

Spec + schema + example: rebuild/docs/build-manifest-spec.md, rebuild/manifest/{build_manifest.schema.json, example-canonical-national.yaml} (branch claude/scoreboard-first-rebuild). Validated against the schema; 31 fast tests green.

Still deferred (keep arch for now): rename to ledger + move the central store to PolicyEngine Supabase — see arch-data#49 and #77.