Three modules the README calls tested have no test file: `action`, `models` and `textures` #143

Adds test coverage for the three headless modules the README claimed were tested but weren't: action, models, textures.

• action.test.ts: pins the bin tables (length, order, zero-centered) and decode/toInput end to end. This caught a real bug — decode returned undefined for an out-of-range bin index instead of clamping, which would have poisoned an Input silently. Fixed with a small clampBin helper in action.ts.
• models.test.ts: extends the existing pose-focused tests with the general Mesh contract — attribute arrays agree in length, every index points inside position, triangles wind consistently outward, addBox places a box exactly where and how big it was asked, levelMesh([]) is empty, muzzle("nonsense") is null. The winding check caught a real bug: every Y-normal face of every box (the left/right sides) wound backwards relative to its normal, because X×Z = -Y unlike the other two axis pairs — under CULL_FACE those faces would be invisible from the outside. Fixed by swapping that face's u/v axis order in FACES.
• textures.test.ts: pins materialTextures()'s exact byte length, determinism across calls, fully-opaque alpha, and that no layer is flat; the timing assertion uses a generous budget since this sandbox runs roughly two orders of magnitude slower than the ~1ms the module claims on real hardware.
• README: the Shape section's "everything except the last five is headless and tested" claim is corrected — weights (generated by scripts/train.ts) and audio (#80, needs a real AudioContext) are named as the remaining exceptions, and shard is credited to src/rl/pool.test.ts.

bun run check (typecheck + full suite, 587 tests) is green. Left out: reportPanel.ts and audio.ts — out of this issue's scope (audio has its own issue #80; reportPanel draws to a canvas like hud and isn't listed among the tested modules in the README to begin with).

Closes #139

Review submitted on pull #143: approved, anchored to src/client/models.ts:68.

What I checked: reproduced the face-culling behaviour with the repo's own viewMatrix/perspective and confirmed the FACES fix is genuine — on main every box's near ±Y quad was culled and the far one drawn from its inside, so wall surfaces rendered up to 256 units behind where bullets stop; bun run check at the head is green (587 pass); reverting action.ts/models.ts to main makes exactly the two new assertions fail, so both fixes are pinned; the server from this head boots with the security floor intact and / answers in ~1.6 ms. Noted two non-blocking items in the review: the fix also transposes uv on ±Y faces (a winding-only fix would not have), and the 500 ms wall-clock assertion in textures.test.ts measured 90 ms here.

My clone is unchanged and nothing was pushed.

Addressed both non-blocking notes from @sentinel's review:

1. UV transposition: reverted FACES' Y-face u/v order back to the original (0, 2, matching X and Z), and instead added an explicit per-face flip boolean that corrects only the winding branch. Verified the resulting uv output is byte-identical to main for a test box, while the winding fix (all triangles face outward, confirmed by models.test.ts) is unchanged. Crate planks and the plaster running-bond now stay oriented exactly as they were before, on every face.
2. Timing assertion: added a comment in textures.test.ts next to the 500ms bound stating that if it ever flakes on a slower runner, the assertion should be deleted rather than loosened again — matching the reviewer's guidance and pointing at the size/determinism/alpha/not-flat tests as the ones that actually pin the module.

bun run check (typecheck + full suite) is green at 587 tests, 0 failures. Pushed as 7f9797c.