D0.5 auto_detect + D0.2 WireTokenAgreement stub (Phase 0, 66/66 tests)

[AdaWorldAPI]

Summary

Two Phase 0 Wire-surface deliverables from the codec-sweep plan, plus board hygiene (STATUS_BOARD + EPIPHANIES) in the same PR per the Mandatory Board-Hygiene Rule.

66/66 cognitive-shader-driver --features serve tests pass (+11 new).

D0.5 — auto_detect (CODING_PRACTICES gap 1 remediation)

crates/cognitive-shader-driver/src/auto_detect.rs (~300 LOC, 8 tests).

Reads <model_path>/config.json (HuggingFace layout) and returns:

pub struct ModelFingerprint {
    pub architecture: String,           // llama / qwen3 / bert / modernbert / ...
    pub hidden_size: u32,
    pub n_layers: u32,
    pub tokenizer_class: String,        // from tokenizer_config.json, best-effort
    pub vocab_size: u32,
    pub default_lane_width: LaneWidth,  // inferred per architecture + torch_dtype
    pub default_distance: Distance,
}

Architecture routing (mirrors Rosetta v2's Python-side heuristic):

Architecture family	Default LaneWidth
llama / qwen / qwen2 / qwen3 / mistral / mixtral	BF16x32 (AMX-ready)
bert / modernbert / xlm-roberta / generic	F32x16 (AVX-512 baseline)
any, with explicit torch_dtype: "bfloat16"	BF16x32 (override wins)

Typed errors: ConfigMissing / Io / Parse / MissingField { field }.

Tests: llama / qwen3-with-tokenizer / bert / modernbert (via architectures fallback) / xlm-roberta (via d_model alias) / generic / missing-config / missing-hidden_size.

D0.2 — WireTokenAgreement stub (the I11 cert gate surface)

crates/cognitive-shader-driver/src/wire.rs (~100 LOC, 3 tests).

Ships the DTOs now; real decode-and-compare harness lands in Phase 2 D2.1–D2.3.

pub enum WireBaseline { Passthrough }  // extensible

pub struct WireTokenAgreement {
    pub model_path: String,
    pub reference: WireBaseline,         // defaults Passthrough
    pub candidate: WireCodecParams,
    pub prompt_set_blob_id: u64,
    pub n_tokens: u32,
}

pub struct WireTokenAgreementResult {
    pub top1_rate: f32,
    pub top5_rate: f32,
    pub divergence_positions: Vec<u32>,
    pub per_layer_mse: Vec<f32>,
    pub candidate_latency_us: u64,
    pub reference_latency_us: u64,
    pub stub: bool,         // Phase 0 honesty flag
    pub backend: String,    // "amx" / "vnni" / "avx512" / "avx2" / "legacy" / "stub"
}

Pass gates when the harness lands (D2.1–D2.3):

• top1_rate ≥ 0.99
• top5_rate ≥ 0.999

This is the actual codec cert gate. Reconstruction ICC is necessary-but-not-sufficient (#219/#220 lesson).

Board hygiene in same commit

• STATUS_BOARD.md — D0.1 Queued → Shipped (was stale, epiphanies: E-ORIG-1..7 + E-MEMB-1..13 + AGI Grundlagenforschung synthesis #227 merged it); D0.2 + D0.5 Queued → In PR.
• EPIPHANIES.md — two PREPEND entries:
  • D0.2 stub flag is anti-codec research: CAM-PQ solves argmax blind spot (ICC 0.9998 at 6 B/row) + production plan #219 defense at the type level. stub:true / backend:"stub" makes "synthetic-rows-mistaken-for-real" machine-checkable, not just documented. Pattern generalises: every Phase-N DTO that lands before its Phase-N+k harness should carry an explicit stub flag.
  • D0.5 auto_detect is the concrete Python↔Rust handshake mechanism. Same architecture→lane-width table in Rosetta v2 (Python) and auto_detect.rs (Rust). E-MEMB-11 handshake moves from conceptual to implemented.

Phase 0 state after this PR

D-id	Deliverable	Status
D0.1	WireCalibrate + WireTensorView	✅ Shipped (#227)
D0.2	WireTokenAgreement stub	✅ This PR
D0.3	WireSweep streaming endpoint	⏳ Queued
D0.4	Surface freeze (rebuild)	⏳ Gates after D0.3
D0.5	auto_detect	✅ This PR
D0.6	CodecParamsBuilder	✅ Shipped (#225)
D0.7	Precision-ladder validation	✅ Shipped (#225)

One more Phase 0 deliverable (D0.3 WireSweep) to freeze the Wire surface.

Test Plan

• cargo test -p lance-graph-contract --lib — 147/147 pass (unchanged)
• cargo test --manifest-path crates/cognitive-shader-driver/Cargo.toml --features serve — 66/66 pass (+11 new)
• cargo test --manifest-path crates/jc/Cargo.toml — 6/6 pass (JC substrate proof unchanged)
• Rules A-F honored: D + E + F apply (D0.5/D0.2 are DTO + serde + one-edge decode)
• Board-hygiene rule: STATUS_BOARD + EPIPHANIES updated in the same scope

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f4304f5bef

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Default missing token-agreement stub marker to true

When WireTokenAgreementResult is deserialized from JSON that omits stub, Serde currently fills it with false, while backend defaults to "stub". That creates a contradictory payload (backend == "stub" but stub == false) and can let clients that gate on !stub treat Phase-0 placeholder metrics as real results. Since this field is intended to prevent exactly that confusion, the default should fail closed (or be required) rather than silently reporting non-stub.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Normalize architectures fallback before family matching

The fallback path copies the first architectures entry verbatim (lowercased) into architecture, but lane-width routing only matches exact family names like "qwen3"/"llama". For configs that omit model_type and use class names such as Qwen3ForCausalLM or LlamaForCausalLM, this yields qwen3forcausallm/llamaforcausallm and incorrectly falls back to F32x16 instead of the BF16 path, so auto-detection returns the wrong default lane width.

Useful? React with 👍 / 👎.