feat(evaluation): repurpose /evaluation as recipe-comparison table

[aistackdev]

Summary

Upstream's /evaluation tab is a gsm8k/mmlu accuracy chart — useful for "did this quantization hurt accuracy" regression work that SemiAnalysis runs. Our team isn't doing that. We're doing recipe selection: given a deployment, which runtime knob combination ships the most throughput at acceptable quality?

So /evaluation is now a recipe-comparison table. Each row is (deployment context) × (technique variant). Within a (model × hw × precision × isl/osl × conc) group, the row whose techniques is empty (or spec_method='none') is the baseline; siblings show speedup = variant.tput / baseline.tput and TPOT-ratio. Accuracy is joined from eval_results at the (model, hw, framework, precision) level since lm-eval runs once per config, not per conc.

What you'll see at https://inference-benchmark.eveningcafe.org/evaluation

Column	What it shows
HW · Precision · Seq · Conc	Deployment context (sortable)
Variant	baseline / MTP×4 / MTP×6 / Eagle / etc. (derived from techniques)
Speedup	variant.tput / baseline.tput — green if >1.01×, red if <0.99×, "—" if no baseline in the group
tput/GPU, TPOT	Raw measurements
Accept Rate	median_acceptance_rate from the variant's metrics (spec-dec only)
Accuracy	em_strict with pp delta vs baseline; "—" if no eval row joined
Run	Link to the GitHub Actions run that produced the row

Default sort: rows grouped together, baseline first, then by speedup descending within group.

Notes

• Tab label renamed Accuracy Evals → Recipe Compare (tab-nav.tsx, landing-page.tsx, tab-meta.ts).
• The old EvaluationContext / ChartDisplay / BarChartD3 components are left in place — they're now orphaned (no imports) and tree-shaken. Cleanup follow-up.
• For our current gemma4 data (no spec_method=none baseline), every row shows Speedup = —. As soon as we ingest a no-MTP baseline run, the speedup column lights up.

Test plan

• pnpm typecheck clean
• pnpm test:unit — 8 new unit tests for buildRecipeRows all pass
• pnpm lint && pnpm fmt clean
• Post-deploy: open /evaluation, confirm gemma4 rows show up with MTP×4 / MTP×6 variants, accept-rate column populated, accuracy column empty (no eval rows for this run yet).

🤖 Generated with Claude Code