Increment 2a: deterministic Hugging Face identity observation for first consumer

[mark-e-deyoung]

Parent: #9
Evidence prerequisite: #12 / draft PR #23
Consumer dependency: #19 / SemperSupra/desktop-ui-cv#17

Objective

Implement only the minimum deterministic normalized Hugging Face observation needed to bind the first desktop-ui-cv KV-Ground consumer to exact upstream evidence.

This increment deliberately de-serializes the broader #9 observatory: it does not wait for Ollama /api/show, hosted-provider selection, all-provider drift/event infrastructure, experiment binding, or standards export before proving the first consumer identity seam.

Input

Captured/sanitized HF fixture + provenance sidecar from #12 for:

vocaela/KV-Ground-8B-BaseGuiOwl1.5-0315

The resolver operates offline in tests. Live provider access remains the capture adapter's responsibility.

Minimal normalized observation

Emit a versioned document containing only:

• provider + upstream locator/requested revision;
• observation timestamp sourced from fixture provenance, not wall clock;
• resolved repository revision;
• explicit identity_strength: repository_commit;
• access state (private, gated, disabled where exposed);
• declared license + source classification;
• sorted file evidence: path, size, Git blob ID, LFS SHA-256/size when exposed;
• raw fixture/provenance references;
• no claim of a whole-artifact content digest unless one is actually computed/proven later.

Critical identity rule

The fixture pair already proves that repository/source revision can move while model payload evidence remains unchanged. Therefore do not collapse source_revision and payload/file identity into one field or one drift conclusion.

Acceptance

• versioned minimal observation schema exists;
• deterministic HF fixture -> observation transformer exists;
• same fixture/provenance produces byte-equivalent canonical JSON output;
• current main resolves to a3c224bdd97ed6de15baef3524eb590c480e0d78 with repository_commit strength;
• exact corrected-weight fixture resolves to fe7563292bb52ab6c235fc3c87157e6a14017479;
• file evidence is sorted deterministically and preserves all four weight LFS SHA-256 identities;
• access and declared-license evidence are explicit without converting license metadata into a redistribution/legal conclusion;
• tests distinguish source-revision movement from payload/file movement;
• no network/model bytes required by tests;
• no Civitai/Ollama/OpenRouter normalization is added in this increment.

WoW / scope

Observe/Discover -> Plan -> Act -> Verify; R4. Prefer a small stdlib implementation. Do not introduce a DB, scheduler, service, plugin framework, global provider ontology, artifact warehouse, or standards export here.

Exit

If this observation cannot provide the exact immutable HF binding needed by the first consumer with a small deterministic transformation, stop and narrow the Foundry MVP before adding more providers.

[mark-e-deyoung]

Verification checkpoint

Exact implementation head: c1694296eeed81a9f5cd11f5dadb13cb945ec976

Public workflow issue-24-hf-observation run 33476469836: PASS.

Verified:

• offline fixture -> normalized HF observation tests pass;
• current KV-Ground main resolves to a3c224bdd97ed6de15baef3524eb590c480e0d78 with identity_strength: repository_commit;
• corrected-weight fixture resolves to fe7563292bb52ab6c235fc3c87157e6a14017479;
• all four LFS weight SHA-256 values are retained as per-file evidence;
• source-revision movement remains separate from file evidence (README differs while tested model/config/tokenizer evidence remains identical);
• declared license remains explicitly declared-metadata-only, not a redistribution/legal conclusion;
• CLI uses captured provenance time, not wall clock;
• absolute checkout prefixes are removed from evidence refs;
• two CLI invocations produce byte-identical canonical JSON (cmp pass);
• tests require no network or model bytes.

This is sufficient for the minimum HF identity binding needed by the first consumer. It does not complete parent #9 or issue #12 and does not authorize acquisition work beyond the separately governed next increment.