Research prompt: sweep-by-workload-class stratification

[Augustas11]

Status

DRAFT — RESEARCH PROMPT ONLY, NO IMPLEMENTATION. This PR contains only a Codex session prompt (specs/RESEARCH_SWEEP_WORKLOAD_CLASS_PROMPT.md) that stages the research + SPEC-drafting work. No code changes.

Why this exists

beta/workloads.py has 6 named workload classes; beta/report.py::summarize_per_workload and beta/DECISION_CRITERIA.md already require per-workload metrics. But the sweep grid is class-blind — 7 contexts × 4 concurrency = 28 cells, no class axis. Autotune winners today are one compromise config across all workload shapes.

Critical timing: SPEC-028 (chain speculative decoding, v0.2 audited on main) is landing now. Draft-token acceptance rate is workload-dependent — reasoning cells gain +49% while code/RAG/long-ctx prefer chain. Without class-aware sweep, SPEC-028 ships with a single --num-draft-tokens N chosen as compromise. Per-class sweep is the multiplier that makes SPEC-028 land at full strength.

What has to achieve to be considered implemented

Ordered gates from research through ship:

1. Research session runs the prompt in this PR and produces:

   • docs/research/sweep-workload-class-2026-07.md — research memo answering questions §A–E with code citations.
   • specs/SPEC-XXX-sweep-workload-class-stratification.md v0.1-draft — normative FRs (search-space shape, winner-picking under SPEC-028, autotune-candidates.json extension, class-parameterized gates), non-goals, open questions, measurable acceptance criteria.
   • .omc/logs/sweep-workload-class-open-questions-2026-07.md — ≤5 items blocked on human decision.

2. SPEC audit — SPEC v0.1-draft passes CODE + SECURITY + ARCHITECT audit lanes at 0-CRITICAL / 0-HIGH (mirrors the SPEC-015 v0.4 lock protocol).

3. Implementation (separate PR, gated on locked SPEC):

   • Sweep search space extended with workload_class dimension (grid-multiplication OR class-filter — SPEC decides).
   • Per-class winners published to phase3-binary/dist/static/autotune-candidates.json (or class-partitioned sibling) with matching signature under the current or new static key.
   • beta/report.py surfaces per-class winner metadata alongside existing aggregate.

4. Measurable acceptance (in the impl PR's tests):

   • Given a prompt from workload class W, provider is served with class-W-tuned knobs (routing mechanism deferred to a follow-up SPEC; this SPEC ships the DATA).
   • Sweep re-run on air5 produces per-class winners that outperform the baseline single-config winner on tok/s or TTFT-p95 for at least 3 of the 6 workload classes.
   • No regression on the classes where the baseline single-config already wins.

Do NOT (until gates 1+2 close)

• Modify sweep code, harness, autotune runtime, or autotune-candidates.json.
• Introduce a request-time workload classifier (called out as follow-up SPEC in the prompt).
• Modify SPEC-013 (CLI autotune) or SPEC-028 (speculative decoding) — this is a companion.

Related

• Companion to SPEC-028 (speculative decoding, merged v0.2 at a1be007).
• Depends on: none.
• Blocked by: none.
• Blocks: none (this multiplier makes SPEC-028 stronger; not required for SPEC-028 to ship).