test(ds4): preserve reference-exact speculative verification

[cheese-cakee]

Summary

• preserve DFLASH_DS4_SPEC_REFERENCE_EXACT=1 as the supported reference-exact speculative verification mode
• keep DFLASH_DS4_FUSED_VERIFY=1 intentionally approximate for throughput experiments
• add a model-backed AR-vs-reference-exact token-ID parity gate with fail-closed evidence checks
• register parser and environment regressions with CTest

This replaces the superseded historical diff on current main. The previous fail-closed fused-flag policy is intentionally not restored: reference-exact verification and approximate fused verification are separate modes.

Exactness contract

The gate holds the target execution policy constant between greedy AR and reference-exact speculation, sanitizes inherited DS4/MoE/MMVQ policy variables, and differs only in DSpark/spec/reference-exact activation. It requires:

• identical complete generated token-ID sequences
• the reference-exact activation banner
• speculative work with exact integer matched / offered counters
• at least one rejection (matched < offered)
• full_snap=1, proving the rollback-capable path was active
• SHA-256 identities for the server binary, target GGUF, DSpark GGUF, and prompt

The gate retains both logs and a manifest instead of overwriting evidence.

Validation completed

• parser/environment unit tests: 11/11 passed
• Python byte compilation passed
• targeted Ruff check and format check passed
• git diff --check passed
• fresh independent final-diff review: no remaining findings

Pending before ready-for-review / exactness claim

• full current-tree CUDA and HIP builds
• model-backed gfx1151 run on the target Lucebox hardware with the pinned artifacts
• observed rejection/rollback and complete token parity from that run

The Lucebox machines and target artifacts are currently unavailable, so this is reopened as a draft. No target-hardware parity or performance result is claimed yet. Historical CI and reviews applied to the superseded implementation and should not be treated as evidence for this rewrite.

Stack

Independent current-main PR. It does not depend on #554 or the #569/#575/#577 performance stack.

[cubic-dev-ai]

All reported issues were addressed across 6 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 7 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 4 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 1 file (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>