reproduce-before-fix

A config plugin — a bug fix must first REPRODUCE the failure as a failing eval, then fix, verify, commit, and file the passing eval. The fail→pass pair on one scenario is the fix's proof (the A/B).

Reproduce before you fix — the fix's proof is a fail→pass pair

If your task is to FIX A BUG, reproduce it first, as a measurement — before you touch the fix. A claim that something is broken is worth nothing until the loss signal shows it broken; a claim that you fixed it is worth nothing until the same signal shows it passing. So a bug fix is bracketed by two readings of ONE scenario:

• A — reproduce (fail). Find the yatsu-core scenario whose expected the bug violates (if none fits, ADD one to the node's yatsu.md — a description + the expected correct behaviour), run it, and file the failing reading with evidence that SHOWS the bug: spex yatsu eval <node> --scenario <s> --fail --note "<what's wrong>" plus an --image/--video of the actual broken behaviour. This is not ceremony — reproducing is how you learn what actually breaks, and a fix aimed at an unreproduced bug aims at a guess.
• B — fix, verify, commit, then file (pass). Four moves, in order: make the code honor the spec; run the SAME scenario against your working tree until it actually passes — confidence is earned here, before any commit; commit the verified tree; only then file the passing reading, spex yatsu eval <node> --scenario <s> --pass, so its codeSha names the very commit you measured (the sidecar append lands last, on top of it). Confidence and anchoring are two different acts: you test the uncommitted tree, but the reading's sha anchor can only land after the commit — this is NOT "commit first, then test" (blind commits with routine rollbacks are bad git). A git sha names only a commit and an uncommitted change has none, so a reading filed from a dirty tree is born mis-anchored — its codeSha is a HEAD that does not contain the change it claims to measure — and freshness later marking it stale is the mis-anchor being correctly exposed, not an engine bug. (A needs no such care: it measures the bug on the old, already-committed HEAD, so its anchor is honest as-is.)

The two readings on the same scenario are the A/B — the error→correct transition, the fix's proof-of-work. yatsu already keeps per-scenario reading history, so the pair is durable and navigable end to end: the dashboard annotator flips between A (fail) and B (pass), and a trunk delivery (deliver-port) carries that pair as the evidence the fix is real. Don't skip A because the fix looks obvious — an obvious fix with no reproduced failure leaves the loss signal blind to exactly the regression you just closed.

This does not apply to building new intent (there is no prior failure to reproduce) — it is the discipline for repair: keep the loss signal honest across a bug's whole lifecycle, not just at the end.