# Challenge gate for Vertex Man

The builder's delivery is a candidate. A separate reviewer receives the frozen source, `BRIEF.md` and `ACCEPTANCE.md`, without the builder's final summary. The reviewer tries to find a material counterexample, then a browser tester attempts to reproduce it. No defect quota: a clean review is a valid outcome. In a repeat comparative run, apply this same gate to **every** first-pass build; the completed run applied the full review and repair gate only to C.

## Reviewer assignment

> Review this game against `BRIEF.md` and `ACCEPTANCE.md`. Look for assumptions that fail when Normal changes to Geometry Inversion, then Time Fracture, then Normal again. Prioritize live objects, collisions, jump direction, target life cycle, controls, UI and restart. Report only material findings. For each, give the relevant code path, an exact player-facing replay, expected behavior, and likely actual behavior. Do not edit the build.

## Browser challenge

Use the same viewport for each candidate. Begin with a fresh restart. Record a target click, a mode change with a live projectile, a jump on the gravity-facing support, a death followed by immediate restart, a resize followed by a shot, and a complete three-mode cycle. Observe the HUD and world together. Repeat restarts ten times. Clear held input while the tab is unfocused and check it on return.

For each result, record **PASS**, **FAIL**, or **NOT_RUN** and the exact input sequence, viewport, expected behavior, actual behavior and console errors. A source hypothesis is not a browser pass. The **Advance update** button is a quick transition probe; the uninterrupted 60/120/180-second sequence is a separate timing probe.

## Repair exit

Copy the frozen source before repair. Give the repair agent only validated findings and the unchanged brief. After repair, replay every failing sequence exactly. Mark a finding fixed only when the observed behavior changes as required. Keep new or untested behavior explicitly open.

## What this run proved

The reviewer found a double-applied canvas offset in C's aiming calculation and a right-gravity support omission. Browser play confirmed a matched shot missed in C and hit in D, and exposed a restart/HUD state leak that the reviewer missed. The exact restart replay passed in D. The natural full-cycle timing and precise live-projectile reversal remain unverified in the browser.