Issue #24: add deterministic Hugging Face identity observation

[mark-e-deyoung]

Purpose

Implement issue #24: the minimum deterministic Hugging Face identity gate required by the first desktop-ui-cv consumer, without waiting for the entire #9 observatory.

This is a stacked draft PR on top of #23 (impl/issue-12-provider-fixtures). Retarget after the prerequisite stack lands.

Included

• schemas/model-observation-v1.schema.json
  • versioned minimal observation contract;
  • provider fixed to Hugging Face for this increment;
  • separate upstream locator/requested revision, resolved source revision, explicit identity strength, access state, declared license metadata, sorted file evidence, and evidence references.
• tools/normalize_hf_observation.py
  • stdlib-only offline fixture -> normalized observation transformer;
  • requires a full 40-hex repository revision and emits identity_strength: repository_commit;
  • preserves per-file Git blob/LFS SHA-256 evidence but does not invent a whole-model content digest;
  • takes observed_at from captured provenance rather than wall clock;
  • canonical sorted JSON output;
  • strips machine-specific checkout prefixes from evidence references.
• tests/test_hf_observation.py
  • exact current-main and corrected-weight revision assertions;
  • sorted file evidence and LFS hash retention;
  • declared-license evidence remains metadata, not a legal conclusion;
  • source-revision movement is not collapsed into payload/file movement;
  • machine-independent evidence refs and byte-deterministic canonical output.
• .github/workflows/issue-24-hf-observation.yml
  • unprivileged public CI;
  • exact-SHA-pinned checkout/setup-python actions;
  • offline tests only;
  • invokes the CLI twice and cmps byte output;
  • verifies repository-relative evidence refs and repository_commit strength.

Deliberate boundary

This increment does not add Civitai/Ollama/OpenRouter normalization, drift persistence, append-only events, experiment binding, standards export, database/service infrastructure, acquisition, or model bytes.

The point is to prove that the already-captured HF evidence can produce the exact immutable binding needed by the first consumer using small deterministic machinery.

Critical identity rule

The underlying #12 fixture pair proves:

• source revision moved from fe7563292bb52ab6c235fc3c87157e6a14017479 to a3c224bdd97ed6de15baef3524eb590c480e0d78;
• model/config/tokenizer evidence remained unchanged;
• README evidence changed.

The normalized observation therefore keeps source_revision and file evidence independent rather than treating repository movement as a payload-digest change.

Pre-commit verification

A local isolated prototype using the exact projected #12 fixtures passed 7 tests before staging. The committed CI adds the repository-relative evidence-ref check and exact-head repeatability proof.

Related: #9, #12, #18, #19, #23, #24; SemperSupra/desktop-ui-cv#17/#19.

[mark-e-deyoung]

Exact-head verification checkpoint

PR head c1694296eeed81a9f5cd11f5dadb13cb945ec976 is verified.

• issue-24-hf-observation run 33476469836: PASS;
  • offline observation tests passed;
  • canonical CLI output repeatability/path-independence proof passed.
• legacy Faster Whisper PoC validation run 33476469791: PASS; publication job did not run on the PR. The unrelated trigger coupling is separately contained by draft PR Contain legacy Faster Whisper PoC CI trigger #26.

This is sufficient evidence for the bounded #24 gate: the reviewed HF fixture can be transformed deterministically into an exact repository-commit-strength observation without network access or model bytes. It does not authorize broader #9 implementation or native acquisition before the first consumer local-handle contract is known.