XPerf — what else (same playbook)

Derived from the closed loop: guide ↔ measured evidence ↔ corpus dogfood ↔ reverse-opt ↔ impl in the target language ↔ honest limits.

Already proven (keep)Copy Anchor Link

• Same-language checkers (Zig / C / JS) + self-lint = 0.
• Corpus reverse-opt improves precision, not app speed by itself.
• Static smell ↔ runtime: partial (jsperf STATIC-VS-V8; cperf STATIC-VS-O2); must re-validate per rule family.
• Cross-kit catalog: smell-catalog.md.
• Engine corpora (gcc/Node) are weak as “find speedups”; strong as FP mills.

High leverage nextCopy Anchor Link

1. Finding → bench gate (jp-delete-shape / jp-delete-array-hole / cp-alloc-hot done)Copy Anchor Link

Every kept rule needs ≥1 fixture pair with JSONL before/after on a named runtime (cc -O2 / Node version / Zig ReleaseFast). Ship as packages/*/evidence/<rule-id>.jsonl + guide citation. Until then, treat lint as review prompt only (already true; make it mechanical).

2. Severity / modes (done in cperf/jsperf)Copy Anchor Link

• error: only rules with measured agreement (e.g. jsperf jp-delete-array-hole, cperf cp-alloc-hot).
• hint / off-by-default: cp-printf-hot, jp-console-hot, jp-alloc-hot.
• Profiles: --preset app|engine|all + --severity RULE=… + --strict-hints.

3. Dynamic companions (not more static rules)Copy Anchor Link

Static answers “might hurt”; dynamic answers “did hurt here”.

4. Own hot code, not only upstream enginesCopy Anchor Link

Point corpora at user/product trees (and mini-program / embed JS for jsperf). Success metric: actionable TPs with benches, not residual count ↓.

5. Cross-kit “same smell, three languages” (catalog landed)Copy Anchor Link

See smell-catalog.md. Fill empty cells (C layout rules) as rules appear. Forces honest portability claims (V8 ≠ QuickJS ≠ C).

6. jvmperf slotCopy Anchor Link

Landed under packages/jvmperf/ (jvmperf-lint, guides, JMH). Evidence gate + correlate: docs/evidence-policy.md.

7. CI dogfoodCopy Anchor Link

make self-lint / npm run self-lint / zig build on PR; corpus lint nightly on pinned SHAs (not every commit).

8. Negative results as first-class docsCopy Anchor Link

Publish negatives as first-class docs — jsperf STATIC-VS-V8 + cperf STATIC-VS-O2 (#if 0 cp-printf-hot) done; keep extending.

Explicit non-goals (for now)Copy Anchor Link

• Full AST/typechecker-grade analysis.
• Claiming engine lint totals = performance wins.
• Detecting JIT tier / IC state from source alone.

Suggested orderCopy Anchor Link

1. Severity split + app preset (c/js) — done.
2. Evidence JSONL for jp-delete-shape / jp-delete-array-hole / cp-alloc-hot — done; see each kit's evidence/.
3. One product/embed corpus instead of more gcc.
4. cperf static vs -O2 — done (correlation suite kept outside the repo).
5. Cross-kit smell catalog stub — done: smell-catalog.md; fill remaining evidence cells.
6. Dynamic companions (perf stat on cperf examples).