[Bench #478] Eliminate generated-artifact drift between canonical results, architecture map and Pages

[godofecht]

Parent: #478

The audit found concrete generated-artifact drift: architecture_performance_map.json can retain older classifications while headline_result_v2.json has newer canonical evidence. Historical SVC predict claims can likewise become stale relative to phase-level current timings.

Make every derived artifact carry and validate the exact source benchmark commit, environment ID and result digest. Architecture maps, optimization roadmaps, README and Pages must be generated from the same frozen canonical input or the local regeneration/validation command must fail.

Hosted GitHub Actions/CI are not part of this work. Validation is local-only.

Acceptance: no derived artifact can silently join an older headline result; the local validator checks source SHA/result digest; stale output is rejected; public Pages displays provenance; one local regeneration command refreshes all dependent artifacts atomically.

[godofecht]

Worked this through as a provenance problem rather than a regeneration-frequency problem.

The smallest robust contract is a shared evidence epoch:

E = (source_commit, environment_id, sha256(canonical_result_bytes))

Every benchmark-derived artifact carries exactly one E. A join is valid only if all participating artifacts have the same epoch and each body matches the digest recorded in one bundle manifest. Static supplements (inventory/profile/roadmap inputs) keep their own digests but do not redefine the canonical epoch.

Under collision resistance, epoch equality is sufficient to rule out silent cross-generation joins: if two admitted artifacts came from different canonical snapshots, equal epochs would require equal canonical content digests. A staged bundle plus one manifest/root pointer also gives atomic publication: consumers observe either the complete old bundle or complete new bundle; mixed files fail the authoritative manifest.

This also clarifies what each field does. The canonical digest is the actual generation identity. source_commit and environment_id are separate audit coordinates rather than substitutes for the digest.

Proposed local-only implementation path:

publish_headline_v2.py creates the epoch from frozen canonical bytes + source SHA + environment; every downstream generator copies it verbatim and declares supplemental-input digests; architecture_performance_report.py refuses mismatched epochs; one local regeneration command stages all derived outputs, writes/validates the manifest, then publishes the set together. Pages displays short source SHA + environment + canonical digest.

I archived the proof, atomic-publication argument, counterexample and an executable Flow provenance witness in godofecht/OpenProjects as ml/content-addressed-benchmark-evidence-epochs.{md,flow}. The Flow model exhaustively checks all 16 old/new assignments across four derived artifacts and accepts only the two uniform bundles.