docs(router): independent replication of the universal-router benchmark; fit_prior Windows fix

[CodeWithJuber]

Summary

An independent replication of the universal router's benchmark (harness-bench run 4) and of the shipped prior, run on a second machine from the pinned public data. Also a one-line Windows fix in bench/universal-router/fit_prior.mjs.

What	Result
Data	SWE-bench/experiments @ 40f164d, sparse blob-less fetch of the 11 runs' per_instance_details.json: 616 KB, under 1 MB in .git
Split	dev 150 / held-out 350: the same issue ids as the original run
Held-out test	217 of 218 metric values identical; only wall-clock fitSeconds differs (37.6 s vs 45.9 s)
Headline	router 76.3% at $0.093 per task vs best-single-on-dev 75.1% at $0.364; cost −$0.271 CI [−0.288, −0.252]; solve +1.1 pt CI [−2.0, +4.3]; pre-registered endpoint met
Prior refit (all 500)	176 of 176 values identical to data/router_prior.json (k=1, scale 4); only provenance.fittedAt differs

The author's stated limit also reproduces: a fixed cascade chosen on dev gets 77.1% at $0.088, level with the router on this fit.

Changes

• bench/universal-router/README.md:
  • the replication table;
  • the missing build routing step (the universal build reads its split);
  • the sparse-fetch commands.
• docs/UNIVERSAL_ROUTING.md: a one-line "Replicated" note under Measured.
• fit_prior.mjs: the default --out used new URL(…).pathname, which is /C:/… on Windows. It now uses fileURLToPath; the path was checked before and after the change.
• Not in this repo: harness-bench's adapters/forgekit_universal.mjs passes plain paths to import() and needed pathToFileURL(...).href on Windows. That file lives in harness-bench.

Verification

• forge docs check passes.
• npm run lint: 0 errors.
• The test suite doesn't cover the bench scripts, and no src/ file changed.

🤖 Generated with Claude Code