[sanskarIN]

Motivation

Matching synthetic fixture hashes and base checkpoint bytes do not establish that base and LoRA generation benchmarks used comparable environments. CPU/CUDA, BF16/FP32, operating system or PyTorch/Transformers version changes can distort apparent latency deltas.

Changes

• Bump local optional Hugging Face FIM smoke protocol to lineloom-hf-synthetic-smoke-v6 and the offline comparison protocol to v5.
• Attach a narrowly allowlisted eight-field runtime profile to opt-in local real smoke summaries: device, precision, operating-system family, Python/PyTorch/Transformers versions, CUDA runtime version and HIP runtime version.
• Reject missing, extra, malformed or source-bearing metadata fields. Require byte-for-byte identical validated profiles on base and adapter runs; no comparison output on mismatches.
• Keep hostnames, full environment variables, serial numbers, checkpoint/model names, paths, raw code and generated completions out of public reports. Output only runtimeProfileComparable and checked field count in source-free comparison summaries.
• Add deterministic offline tests for missing fields, types, changed runtimes, CPU-vs-BF16 rejection and matching reviewed CUDA metadata.
• Update model/train/README.md and model/EVALUATION_PROTOCOL.md with v6 compatibility, privacy and measured-performance limitations.

Validation and boundaries

• Run locked Windows/macOS/Linux CI, independent Linux extension-host test, dependency-security audit, and VSIX install/upgrade checks before merge.
• No GPU experiment, checkpoint download, training corpus collection, new model, live inference, new extension release or marketplace upload is performed.
• Matching profile is not verification of identical physical GPU, drivers, temperatures, process load, execution order, code correctness or usable editor latency. All real-model evidence remains outstanding.