feat(compute): MVP-A inference execution contract (mbongo-infer:v1, kaayu-cv-extract:v1)

[gkalombo21]

MVP-A — AI Inference Execution Contract (mbongo-infer:v1)

First implementation gate on the frozen MVP v0.1 path (docs/product/MVP_v0.1.md §11, #148). Baseline 38445ef278894ec31ac30b12660c6afe5d738f62. Tracks #150.

What it freezes. Exactly what a ComputeTask means when it asks for the MVP inference, without touching the v0.4 protocol:

• execution_spec = "mbongo-infer:v1" (15 bytes) ‖ job_spec_digest (32 bytes) — fixed 47 bytes inside the 1,024-byte RFC 0005 bound; family prefix, then version, then length are checked in that order.
• job_spec_digest = BLAKE3("mbongo:infer-job-spec:v1" ‖ canonical job spec); constrained canonical JSON (sorted keys, no whitespace, integers only, defined escapes; no floats, booleans or nulls) shared by job spec, input object and result.
• InferenceJobSpecV1: runtime llama.cpp / llama-server v0.7.0 @ 620d2225 (MIT, verified at that commit); model Qwen3-4B-Instruct-2507 GGUF Q4_K_M, artifact_sha256 = 3605803b…e597 from the Hugging Face LFS metadata of unsloth/Qwen3-4B-Instruct-2507-GGUF @ a06e946b (Apache-2.0), to be recomputed from the downloaded bytes in MVP-B set-up — not fabricated, not downloaded here; the digest is authoritative, names are not; the GGUF embeds the tokenizer and chat template, so no separate tokenizer id.
• Profile kaayu-cv-extract:v1: input {"contract":"kaayu-cv-extract-input:v1","cv_text":…} ≤ 10,000 bytes; prompt = template file verbatim as the system message + CV text as the user message (template SHA-256 pinned in the job spec); output kaayu-cv-profile:v1 JSON Schema (identity-free: no name or contact field exists; additionalProperties:false, enums, bounds; only keywords llama.cpp's grammar converter supports), schema SHA-256 pinned; generation temperature 0 / top_p 1.0 / seed 42 / max_tokens 400 with frozen ranges.
• Result rule: raw output → strict parse → schema validation → canonical bytes → persisted → output_commitment. Whitespace and key order never change a commitment.
• Worker gate preflight (8 ordered steps, fail-closed, no execution/result/receipt on any failure) and finish; 17 application error codes; retry keeps the job spec immutable by construction; model/runtime drift fails closed at start-up and per task.

Evidence. Named group inference_contract: 25 tests (A01–A24 + negative controls: digest-blind, version-blind, model-digest-blind each accept what the strict gate refuses), golden vector byte-exact (A20) and the committed job-spec.json identical to the canonical bytes; wire shapes of ComputeTask and Receipt asserted unchanged (A18, A19). Full validation in the PR thread. 62 protected paths (RFC 0005, RFC 0002, RECEIPT_SPEC, lock v0.4, rpc_v0.3, consensus, SDK wire/source, reference worker core, conformance, J, K, SNP adapter, protocol vectors, MVP_v0.1.md) byte-identical to the baseline.

Not in this PR. No llama.cpp execution, HTTP front, CLI, GPU, pricing, metering, MBO, SDK change (helpers deferred to MVP-C, which must reproduce the golden vector).

Refs #150, #148, #144, #52.

🤖 Generated with Claude Code

[gkalombo21]

Validation at source 122bce1b28e50910df17f0fce38fc8a49281f134 (local, Windows; CI binds the same SHA)

Check	Result
cargo fmt --all -- --check	clean
cargo clippy --workspace --all-targets --all-features -- -D warnings	clean (exit 0)
cargo test --workspace	all suites ok, 0 failed (ignored: 2 baseline + the SEV-SNP live test)
Mbongo MVP Inference Contract (inference_contract)	25 passed / 0 failed / 0 ignored — A01–A24 + negative controls (digest-blind, version-blind, model-digest-blind caught)
Mbongo Compute Provider Policy	41/41
Mbongo Compute Conformance	PASS 38/38
Mbongo Compute Confidential Authorization	CONFIDENTIAL AUTHORIZATION: PASS (6/6 test group)
Mbongo Compute Hardware Attestation	18 passed / 1 ignored (live test, never counted as pass)
reference harness	REFERENCE COMPUTE FLOW: PASS
replay harness	DETERMINISTIC REPLAY: PASS
SDK	153 tests, 153 pass, 0 fail (no SDK source touched)
docs	79 relative links and anchors resolve across the contract document, INDEX, vectors README and the scope document; 67 section citations; typos (CI's checker, 1.49.0) clean on every touched path; git diff --check clean; 0 CR bytes in any touched file; ci.yml parses (js-yaml)
golden vector	job_spec_digest = be9ae960…fe0e; execution_spec = 6d626f6e676f2d696e6665723a7631 ‖ digest (47 bytes); committed job-spec.json is byte-identical to the canonical bytes (A20)
protected surfaces	62 paths (RFC 0005, RFC 0002, RECEIPT_SPEC, PROTOCOL_LOCK_v0.4, rpc_v0.3, consensus and node sources, SDK wire and sources, reference worker core, conformance, J, K, SNP adapter, protocol vectors, SNP fixtures, MVP_v0.1.md) resolve to the same blob ids at 38445ef2 and at this source: expected 62, found 62, identical

Change surface: .gitattributes (+2), ci.yml (+6, one named step), lib.rs (+1 module line), docs/INDEX.md (+1 row), and new files only — the inference module (5 files), the test group, the contract document, three profile artifacts, six vector files. Nothing deleted, nothing else modified.

[gkalombo21]

Addendum — source 6ec31ab310c6708c68c5dec674d7a1f7eb9a21e1

The first source (122bce1b) failed Docker Devnet Bootstrap only: the devnet image copies Cargo.* and crates/ alone, so include_str! of files under docs/ could not resolve inside the image (every other job and the new named step passed: Rust Checks green, inference_contract 25/25). The three profile artifacts moved to crates/mbongo-compute/profiles/kaayu-cv-extract-v1/ with identical blob ids (fb185b39 job-spec.json, 0ae00839 output-schema.json, e279df66 prompt-template.txt), so the digests and the golden vector are unchanged; every reference follows. Local validation re-run at this source: fmt clean, workspace clippy -D warnings clean, workspace tests 0 failed, inference_contract 25/25, provider policy 41/41, hardware attestation 18/0/1 ignored, conformance 38/38, confidential authorization PASS, reference and replay harnesses PASS, SDK 153/153; 62 protected blobs identical to 38445ef2. Diff against the baseline: 20 files, +3861 / −0.