fix(release): prove independent distribution rebuild reproducibility before publication

[seonghobae]

Release acceptance gap

Protected main@493326f2de49ea1704da0ded19868ed05d2fe00f requires more than artifact existence and provenance. docs/PRD.md PRD-FR-081 requires release evidence for packaging, provenance/SBOM and reproducibility; docs/TRD.md likewise lists release artifact digests/SBOM/provenance and reproducibility evidence as separate acceptance inputs.

The current protected release-evidence index computes SHA-256 over the one set of wheel/sdist files it receives. That proves the identity of those files, not that a clean independent rebuild of the exact same source/toolchain reproduces them. Active supply-chain writer #1692 correctly owns SBOM/provenance and immutable draft→asset/PyPI→publish transaction safety, but its current scope and issue #1691 do not claim independent rebuild reproducibility.

This is therefore a successor release-reproducibility lane, not an expansion of #1692.

Single-writer and sequencing boundary

• Do not modify .github/workflows/publish-pypi.yml, .github/workflows/release-tag.yml, or their contract tests while fix(release): publish SBOM and build provenance #1692 is open. fix(release): publish SBOM and build provenance #1692 remains the canonical writer for the current publication/SBOM/provenance transaction.
• Do not compete with chore(release): cut v0.9.2 #1471 for version/changelog/tag-cut authority.
• Start implementation only from the protected head after the relevant release-workflow prerequisites land; refetch intervening deltas and adopt/adapt them with a non-force restack.
• This issue owns only reproducible package-build evidence for fast-mlsirm release artifacts. It does not own psychometric numerical determinism (for example fix(rotation): make Oblimax deterministic-reference math explicit and reproducible #1747), TEPP temporal/event composition, or contextual-orchestrator routing.

Scientific/engineering claim boundary

Distinguish these claims explicitly:

1. Artifact identity — SHA-256 identifies one produced file. Existing release-evidence indexing already does this.
2. Same-environment clean rebuild repeatability — two clean builds under the same locked target/toolchain produce byte-identical output.
3. Independent-runner reproducibility — independent clean runners for the same target/toolchain produce byte-identical output.
4. Cross-target scientific/runtime parity — different OS/architecture wheels need not have identical bytes; their numerical/API parity is a separate fast-mlsirm contract.

Do not describe (1) as evidence for (2) or (3), and do not require byte identity across different wheel target triples.

RED → GREEN contract

1. Add focused fail-first release-contract tests showing publication cannot finalize when the required reproducibility manifest is absent, incomplete, source-mismatched, target-mismatched, or reports unequal digests for a same-target rebuild pair.
2. Derive the build time input from the exact release source commit and export a documented SOURCE_DATE_EPOCH rather than wall-clock time. The Reproducible Builds specification defines SOURCE_DATE_EPOCH as the standardized build-time input for reproducible output; use the exact commit timestamp and record it in evidence.
3. For the sdist and for every published wheel matrix target, perform two clean builds from the same exact release_commit, immutable dependency/toolchain inputs, target triple/interpreter contract, and build options. Do not reuse target/, dist/, wheel caches, or a predecessor artifact as the second build input.
4. Prefer independent jobs/runners for the two members of each comparison. If a target cannot yet provide independent-runner equality, fail closed for the stronger claim and record the narrower verified level rather than silently downgrading acceptance.
5. Compare complete artifact bytes by SHA-256 after canonical filename/target matching. Fail on missing, duplicate, extra, renamed, or many-to-one matches; do not normalize or rewrite the generated artifact to manufacture equality.
6. Emit one machine-readable reproducibility manifest bound to release_commit, version, commit timestamp / SOURCE_DATE_EPOCH, Cargo.lock hash, Rust toolchain, maturin version/action SHA, runner image identity, target/interpreter tag, build arguments, both artifact digests, and comparison result.
7. Feed that manifest into the existing release-evidence index and release acceptance. The index must distinguish artifact_digest from rebuild_reproducibility; a single original artifact digest cannot satisfy the new requirement.
8. Attach/attest the accepted reproducibility manifest through the canonical release transaction after fix(release): publish SBOM and build provenance #1692 lands. Preserve builder-local provenance; the comparison/aggregation job must not masquerade as the builder of either distribution.
9. Keep publication fail closed: GitHub Release finalization and PyPI publication must not proceed when the required same-target reproducibility evidence is absent or failed.
10. Preserve token-based PyPI semantics and all branch/review gates. No force push, gate weakening, source rewriting, skipped/xfail acceptance, cache-only second builds, or denominator tricks.

Required evidence

For each artifact target record:

• exact source SHA and source commit timestamp;
• clean-build A and B runner/toolchain identities;
• dependency-lock and build-action identities;
• wheel/sdist filename and target tag;
• SHA-256(A), SHA-256(B), and exact equality verdict;
• failure class when equality is not achieved;
• retained artifacts/logs sufficient to diagnose timestamp, archive-order, generated metadata, linker, toolchain, or platform-image nondeterminism.

Release documentation must state exactly which reproducibility level is proven. A same-runner or same-image double build must not be advertised as universal cross-platform bit reproducibility.

Acceptance

• Focused RED demonstrates that current protected main can produce/hash one artifact set without proving a clean rebuild.
• GREEN demonstrates byte-identical same-target rebuild evidence at the declared level for the complete release matrix, including sdist, without shrinking the matrix.
• Exact-current CI/security/coverage/review gates pass and an independent current-head approval exists.
• docs/release_acceptance.md, TRD/TEST_STRATEGY/OPERABILITY/recovery/release evidence docs and the governed changelog describe the claim, failure semantics, rollback, and diagnostic path without overstating cross-platform guarantees.
• docs/product-technical-gap-baseline.md is updated only through its canonical single-writer lane.

Basis

• fast-mlsirm PRD-FR-081: exact integrated protected-head release requires packaging, provenance/SBOM, reproducibility, compatibility, review, rollback/migration, and release-acceptance evidence.
• fast-mlsirm TRD release acceptance separately requires artifact digests/SBOM/provenance and reproducibility evidence.
• Reproducible Builds, SOURCE_DATE_EPOCH: https://reproducible-builds.org/docs/source-date-epoch/

This issue remains open until the reproducibility evidence is integrated on protected main and exercised by an actual release transaction; a single successful local or synthetic test is not release completion.