Unaccepted export/animation studies. No playable gunpack or native performance claim.

This combines the corrected v32-r3 rear contours with the existing three-channel recoil fixture. The decorative parts follow the weapon root, their intended scale is retained, and the original four-projectile gameplay burst remains a separate contract.

Middle and lower channel samples

• 340 integer-frame checks cover 6,120 owned-part transforms, with zero observed owner-follow or inactive-motion error.
• 119,880 quarter-frame moving-versus-other triangle-surface comparisons report no baseline or new intersections in this scene. Same-owner pieces, containment, actual hands, attachments and continuous/native interpolation remain outside this result.
• Three position clips contain 111 keys total; exported JSON is byte-identical to the data accepted by the isolated released TaCZ parser on Java 17/21. Bone binding, material glow, gameplay and actual playback remain untested.
• A clean extraction rebuilt the combined scene, reproduced all four scoped check reports apart from scene-file hashes, and produced byte-identical animation JSON.

Small v33 build/check patch · Detailed evidence and limits. Build from the exact v31 input archive listed in the patch README. The patch includes the corrected root-ownership stage, avoiding another full-scene backup.

Still unaccepted: final concept fidelity, exportable textures, player hands, full animation/effect set, native TaCZ/TML binding, multiplayer, optional integrations and measured performance. These authoring checks do not close those gates.

This export is tied to the published v15 source, not the later receiver study. It is unaccepted for native use.

The corrected per-owner bake produces 1024×1024 base-color, emission-color and linear roughness maps. 142 separately owned parts preserve 113,160 UV corners: position and UV errors are zero; maximum corner-normal vector error is 0.00033557; the controlled owner-translation check is zero. The roughness PNG has equal RGB channels and is treated as non-color data. Its alpha is not a reliable geometry coverage mask.

The authoring chamber animation is frozen at the source frame in this atlas. Linked bump/normal effects still depend on authoring nodes; they are not exported normal maps. Native material/shader support and animated energy remain separate work. Do not present this as a completed PBR or animated texture pipeline.

The 64-sample matched render has whole-image mean absolute RGB error 0.001382, weapon crop 0.003750 and chamber crop 0.016141 on a 0–1 scale. The much larger local chamber error exposes blur that a background-heavy whole-image average would hide. Render noise is included; these numbers are not acceptance thresholds.

No sampled base-texel source-owner overlap was found. Simple 16×16 box footprints mix owners in 598 occupied blocks. This is conditional mip risk, not measured native bleeding; dilation, same-owner islands and bilinear/trilinear/anisotropic sampling still need investigation.

Rebuild with Blender 4.3.2: run bake.py, check_parts.py, then render_comparison.py. The sibling discharge-finish-study-v15 source comes from the earlier published authoring ZIP. The published atlas-mip ownership checker can reproduce mip-ownership.json. Neither native rendering, hand poses, collision, complete animation nor performance is certified.

Unaccepted authoring/transport fixture based on the published v 15 exterior model, not a playable gunpack.

18 authored pieces are assigned to three independently owned channel groups, top to bottom. Each has a short recoil/recovery curve and its own glow material. The fixture uses shot starts at frames 1,166,331 at 60 fps, reflecting the baseline's nominal 55 ticks/20 TPS interval for illustration. It does not schedule game firing. Each accepted original burst still has four projectiles; channel count is not pellet count.

check.py evaluates 340 integer frames and 6,120 part transforms: maximum owner-follow/inactive-motion error 0, all three glow materials return to neutral outside their active window. This is not continuous collision, interpolation, hand or play-feel acceptance.

export_clips.py emits three position-only clips with 37 keys each,0.15 seconds long. Blender-X maps to model+Z at 4 model units per authoring unit. Maximum linear sample error over 361 points/clip is 0.003077 model units. Glow/shader keys are not exported in these clips.

The actual released TaCZ 1.1.8-hotfix 2 animation-parser bytecode accepted all 3 clips/111 keys onJava 17 andJava 21. The harness validates 18 hash-pinned released classes and five pinned library JARs; documented narrow Minecraft/logging helpers are stand-ins. This does not test a real game, native animation playback, model bones, resource binding, audio, multiplayer or performance.

To reproduce authoring checks, run Blender 4.3.2 with --python-exit-code 1 --python build.py, then check.py and export_clips.py. check_released_parser.py expects the separately documented released-tacz-parser-study cache and local Java 17/21 tools; it does not download them. Do not put these sources in a Mods folder.

A future native adapter must resolve matching bones once per asset generation, select only the confirmed channel, reconcile local prediction without adding network latency, preserve original resources/cadence and restore interrupted poses. The generic confirmed-shot component is published separately.

Source, maps, editable scenes and tests · Required v15 authoring source · Released parser harness · Confirmed-shot presentation component

Partial authoring/export study. Native weapon and final-art acceptance remain open.

Open the optional static 3D inspection page · Editable Blender/transport package · Detailed evidence.

The revised atlas supplies color, emission, roughness and tangent-space normal maps. The earlier bake was rejected for overlapping/collapsed UV faces and a corner-normal mismatch. Triangulation, unwrapping and isolated sliver repairs now pass the documented UV tests. Small islands, mip behavior and final texture detail still need attention; passing overlap checks does not finish the artwork.

Bounded check	Result
Revised atlas	75,188 triangles; zero degenerate UV triangles and sampled overlaps
Continuous UV triangle intersections	Zero positive intersections above 1e-14 UV-area tolerance
Grouping	186 authoring parts → seven equal-motion groups
Quarter-frame position checks	52,346,400 vertex samples; maximum error about 1.09e-7 scene units
Preserved attributes	Exact UVs and material assignments; normal-vector error below 0.001
Static GLB	Seven meshes, 21 primitives, four images; no animation export
Translation clips	Three clips, two moving groups each, 222 total keys
Compact transport JSON	17,180,376 → 12,608,390 bytes; consumed float32 arrays unchanged

The released TML data baker parses the geometry on Java 17/21, but its flat-normal diagnostic finds eight faces with normal length below 0.99. This is an unresolved lighting concern, not an accepted result. The separately proposed authored-normal mode preserves all 300,752 expanded corners in isolated source tests and leaves default arrays unchanged. Stock TML does not implement that opt-in proposal. Native fallback/VBO parity, material maps and shader integration still need real game tests.

The static GLB roundtrip closely matches the reviewed authoring image; it is not the Minecraft renderer. The normal bake changes some procedural shading, and the chamber pattern is frozen at one source frame. Native animated energy, sound, hands, confirmed-shot event binding, complete movement/reload/inspection states and measured performance are not supplied by this study.

The source concept remains more sculpted and ornamented than the current model. Grip contours, receiver detail, crown and lower-rail shapes, metal ornamentation, final surface art, variants and the full effects package remain explicit work. Seven groups and smaller JSON do not establish an FPS improvement or final fidelity.

Authoring/export progress only. Final concept fidelity and native acceptance remain open.

Editable scene and reproducible export package · Static GLB · Scoped evidence.

The first v37 grip bake blurred its fine pattern into waves and failed visual QA. A dedicated grip atlas corrected that, but added two images. v39 instead keeps continuous grip UV charts and allocates more useful texels within the original four 2048² maps. It incorporates the refined grip and v38 chamber collars. No extra detail atlas or second UV set is needed in this authoring export.

Check	v39 result
Geometry	78,020 authoring triangles, 186 parts grouped into seven meshes
Grouping samples	54,261,280 vertex samples at quarter-frame steps
Maximum sampled world-position error	8.34e−8 scene units
UV/material preservation	Exact grouped UVs and material ownership; normal-vector error below 0.001
Atlas geometry checks	No degenerate UV triangles or positive-area overlaps under documented tolerances
Texture caveat	Tiny triangles without a texel-center sample remain; these checks do not certify every mip level
Static GLB	7,848,188 bytes, seven meshes, 21 primitives, 12 materials and four images
Transport	17,537,932 → 12,898,214 bytes, 26.46% less JSON
Released baker equivalence	2,496,640 float values bit-identical on Java 17 and 21
Clean package check	Geometry, animation JSON and GLB regenerate byte-identically from the packed grouped scene

Matched Blender renders were visually reviewed. Whole-frame source-versus-baked mean absolute RGB difference was about 0.00295 on a 0–1 scale; grip baked-versus-GLB difference was about 0.000170. These are controlled render comparisons, not perceptual proof, native screenshots or performance results.

The stock released TML baker still reports eight short flat normals. The separately proposed authored-normal path passes the explicit v39 fixture for all 312,080 expanded corners, but remains a source-only experiment. Native material routing, fallback/VBO lighting parity, animated energy, hands, LODs and frame-time measurements are unresolved.

The package reproduces exports from its included baked scene. It does not claim to reproduce UV preparation and the full bake from scratch. The existing v35 inspection page remains labeled v35; it must not be mistaken for this newer model. Most importantly, the target concept is still more sculpted and ornate: this improved texture/export result does not close the final-art gate.