Issue #868: benches/overlap/lowered.sh writes the manifest key ordering, which the manifest refuses: the bench measures nothing

Open and answer this issue on GitHub.

ready · opened 2026-10-26 · updated 2026-10-26 · 1 comment


agent: builder model: sonnet max_turns: 80 budget_usd: 15 — Found by the #645 builder (report docs/briefs/b645-overlap-shared-mut.md.step2.report.md, commit 1702057).

Facts

benches/overlap/lowered.sh builds benches/overlap/pair.nika twice, writing the manifest key ordering = "effects" / "strict" (lines 38-49) and grepping task::read_pair in the generated Rust. The manifest no longer takes ordering (Part III, docs/specification/30-nikaia-tooling.md), so the build fails and the bench measures nothing. benches/overlap/README.md says so since 1702057.

To build

  1. Rewrite the bench as benches/overlap/lowered.nika (CLAUDE.md: helper tools are Nikaia), run with nikaia run: one source, two builds that differ only in whether the pair is lowered onto task::read_pair (find the switch that replaced ordering: user-parallelism in Part III’s manifest table). Output the same table as before; counts, not times (CLAUDE.md, Measuring performance: callgrind instructions).
  2. Delete lowered.sh, update the README’s method and table row. A missing std entry on the way is a finding (file it), not a detour.

Not a version raise (benches/).


Generated by Claude Code

Comments

keywan-ghadami, 2026-10-26

Architect, after 602904c: the dead ordering key is gone since 356a2e3 (lowered.sh now builds with user-parallelism and checks task::overlap2), so the bench measures again. What is left is step 1 only as the CLAUDE.md principle (move the shell script to lowered.nika); low priority, label stays ready.


Generated by Claude Code

Back to Open Issues