[sanskarIN]

Motivation

Raw base-versus-LoRA generation latency is misleading when the two models emit different output lengths. The optional local HF smoke currently reports timing but not actual generated-token counts, so a slower model could simply be generating more tokens.

Changes

• Bump local HF synthetic smoke to lineloom-hf-synthetic-smoke-v5 and comparison to lineloom-hf-synthetic-comparison-v4.
• Count actual new token IDs (including generated stop IDs), immediately after local model.generate(); never persist decoded code, token IDs, per-case lengths, prompts or paths.
• Aggregate a bounded generatedTokenCount and five-language generatedTokensByLanguage, alongside existing p50/p95 generation-only timings.
• Make the offline report comparator validate mandatory token accounting, per-language attempt/length bounds and total consistency, rejecting previous protocols or fabricated counts.
• Expose source-free base/adapter and per-language token deltas and explicit statuses DIFFERENT_GENERATED_TOKEN_TOTALS / AGGREGATE_TOKEN_TOTALS_MATCH_ONLY. Even equal aggregate counts do not prove per-case length parity or hardware parity.
• Extend pure offline Python tests and document all limitations in model/EVALUATION_PROTOCOL.md and model/train/README.md.

Safety and release boundaries

• No checkpoint downloaded, model trained, actual GPU inference run, licensed dataset collected, new local model weights published, or editor inference behavior changed by this PR.
• Real smoke remains explicit opt-in with locally provisioned reviewed checkpoint and adapter.
• Merge only after full cross-platform tests, security checks, Linux mocked extension-host and VSIX candidate install/upgrade checks. No marketplace/registry release or publisher migration.