0.3.4：修掉让「MC 模组工作室」模式整轮跑不起来的 preset 变量

dsh-mc-art-panel 0.3.4

0.3.4

Fixed — one preset variable broke every session on the mode

presets/mc-studio/agent.cordis.yml carried suffix: Your working directory is {{cwd}}. in the
persona, and named the same variable again inside the prefix body. The current runtime registers only
provider and model for that prompt section, so rendering the system prompt threw
unknown prompt variable "{{cwd}}" and the whole turn failed before the model ever ran — the
session showed nothing but "处理失败 · 本轮运行失败".

• The working-directory sentence is gone and the prefix now says 本会话的工作目录; no prompt text
  in the preset references an unregistered variable.
• tools/gen_preset_patch.mjs --write regenerated both the standalone desktop-generation.patch.yml
  and the copy embedded in panel/cordis.patch.yml, so the shipped package carries the fix.
• New gate preset-vars-test.js (plugin-preset-vars): every {{var}} in the preset YAMLs and
  the packaged patch must be within {provider, model}; --fault injects an unknown variable and
  exits 1.

Note

An agent's composition is fixed when its session is created, so a session that already failed keeps
failing. Restart DSH, then start a new conversation on the mode.

0.3.3：设置视图按专业工具重做；修掉「网格线透过方块」的假透明

dsh-mc-art-panel 0.3.3

0.3.3

Round one of the UI review targeted the settings view, because that is where "AI-made" showed
most: a flat pile of rows with inconsistent label columns, no grouping, and a typo.

Changed — settings, reviewed from a screenshot

• It is a real property grid now: a fixed 88px label column with labels right-aligned, values
  starting on one axis, hints under the value column. The 路径 row no longer carries its own
  indent (~39px off from every other label).
• Grouped: 参考目录 / 生成 / 命名空间 / 通知 / 文件 — each an 11px/600 muted header with
  a 1px hairline, instead of one undifferentiated list.
• A view, not an insert: settings opens as its own page with a 返回资产列表 control; previously
  it was stacked above the asset list and squeezed both the list and the 3D viewport.
• Wording bug fixed: the detection line said 2242 张截图 ("2242 screenshots") where it meant
  textures — now 2242 张贴图, with the number in monospace and tabular figures.
• Long paths (设置文件 …) were clipped by the panel edge without an ellipsis; they truncate
  properly now and carry the full path in title.
• Buttons share one 22px height; checkboxes are 12×12 with a 1px border and a brand-coloured check
  when on, plus :focus-visible.

Fixed — the ground grid no longer shows through the block

The 3D viewport drew its ground grid after putImageData, with canvas 2D strokes. That paints
over every pixel already rendered, so grid lines crossed the block's faces — which reads exactly like
the block being semi-transparent. It never was: the rasteriser rejects faces below alpha 0.5 and
writes depth without blending, so the faces are opaque.

The grid is now rasterised into the colour/depth buffer with a depth test, so the block occludes
it and lines only appear on background pixels. The pick buffer is still untouched (the grid is not
clickable). A new gate, grid-occlusion-test.js, executes the emitted function against two synthetic
buffers and asserts that not one pixel inside a nearer occluder region gets stained — and
--fault deletes the depth test to prove that assertion can fail.

Gates

• ui-test gained behaviour assertions for this view: the gear opens it, it is not rendered on
  the same screen as the list/viewport, and 返回资产列表 brings the list back — plus
  --fault-settings-back, which deletes that button's handler and turns the assertions red.
• ui-style-test gained assertions for the property grid, the group headers and the checkbox.
• 51 conditions, two passes (default console encoding and --encoding gbk): 50 green / 1 SKIP /
  0 red.

0.3.2：UI 按书面规范重做第二轮（第一次是看着截图改的）；修掉「宿主已卸载」与落点竞态

dsh-mc-art-panel 0.3.2

0.3.2

The UI pass this project needed was never "make it nicer": it was look at it. This release starts
with a written spec, then a machine that can see the result.

Added — two documents

• docs/UI-STYLE.md — a design spec, not a mood: the reference language is a professional
  inspector (Blender properties / VS Code panel — flat, dense, narrow, quiet), with an explicit
  anti-list of what made the panel read as "AI-made": cards inside cards, large radii, pill buttons,
  gradients, glow, animation, coloured chips, emoji icons, 16px+ padding, centred form labels.
  It fixes the numbers: 12/11px type, 22px rows, 26px form rows, 28px toolbar, 22px controls,
  radii 2/3 only, spacing 2/4/6/8, four interaction states, token-only colours.
• docs/CUA-UI-REVIEW.md — how an agent can see this panel: the official
  @deepseek-ai/dsh-experimental-computer-use-cua-driver-mcp provider (published by DeepSeek,
  packages/experimental/… in deepseek-ai/deepseek-harness) mounted against the locally installed
  Cua driver, plus the exact screenshot/zoom/click procedure and its traps (Electron exposes no UIA
  tree; Chromium input fields only accept foreground delivery; coordinates are in the returned
  screenshot's pixel space).

Changed — the panel, judged from screenshots

• Header: the breadcrumb leaked placement jargon (中央面板) and repeated the tab title. It is
  now one row — path only (mono 11px, muted, ellipsised, full path in title).
• Asset list: rows are two columns — localised name on the left, monospaced id right-aligned
  and muted — instead of 深渊砖 abyss_bricks run together. The group header (方块 · 44) is
  sticky with a hairline, the list fills its container and scrolls on its own, and the selected row's
  2px left bar uses a brand-derived token (--mc-brand-bar, brand 75% + text 25%) rather than
  reading as grey.
• Toolbar: counts (6 面 · 贴图 1/1) in mono muted on the left; 重置/手动修改 as icon buttons
  with tooltips and 1px dividers; @ 提意见 as the only primary action, right-aligned.
• 3D viewport: the camera now auto-fits the bounding box by projecting all eight corners
  into an 88% safe box (~12% margin), reserves room for the hotbar and shifts the content up so the
  block is never clipped, and re-fits on asset switch and viewport resize. A subtle ground grid and
  a small axis indicator give the model a ground and an orientation; the selection label moved from
  a chip floating over the hotbar to a mono label in the corner.
• Narrow widths: the settings "使用" button was clipped by the right sidebar; the input and button
  group now wrap instead of overflowing.

Fixed — recovery instead of a raw error

• "宿主已卸载" is now actionable. After repeated rebuilds plus HMR the host half can end up
  disposed (0 个方法), and the panel used to show that sentence and nothing else. It now says, in
  two lines, that the plugin host did not load (which happens right after a rebuild) and that DSH
  must be fully quit and reopened — with the technical detail collapsed under 详情 — and it
  retries the operation once after a second before reporting.
• Panel placement is deterministic. At boot the right sidebar service can appear after the panel;
  the panel now retries the migration and removes the left/centre fallback once it lands on the right
  sidebar. ui-test --fault-placement proves the retry is load-bearing (delete it → three assertions
  go red).

Gates

• New tools/mcart-plugin/ui-style-test.js (plugin-ui-style): 27 assertions over the CSS string
  that actually ships — every anti-list item must be absent, every fixed number present; three fault
  modes return exit 1. icon-test gained static assertions for fit/reset/grid/axis/details/retry/
  two-column rows/placement retry; ui-test gained the placement-race and host-unloaded behaviours
  without relaxing any existing behaviour assertion.
• 51 conditions, two passes (default console encoding and --encoding gbk): 50 green / 1 SKIP /
  0 red.

Note

The visual review was done on the previous build with CUA screenshots; this batch's final look was
checked by code, tests and gates, and needs one page refresh to be seen.

0.3.1：@提意见直接插进输入框；三个「报告在指责产物」的坑修掉；母模型表单一真相

dsh-mc-art-panel 0.3.1

0.3.1

Changed

• 「@ 提意见」inserts straight into the composer. It used to set a pending reference and open a
  bar ("引用给 AI" / path / 插入输入框 / 取消) that needed a second click. It now appends @<path>
  (plus the sentence describing a settings change) to the draft immediately, keeping what the user
  had already typed and leaving a trailing space. The dock entry stays registered because its render
  is the only place the shell hands over inputActions.setDraft — it renders nothing on the normal
  path and is kept solely as the cold-start fallback: if the handle is not available yet the bar
  appears once, so a click never silently does nothing. Settings' "tell the AI" reuses the same
  insertion path, and the reference bar now says so when the scene has no file path instead of
  ignoring the click.
• The red-line gate was retargeted to the new insertion point, so "insert writes the draft and never
  a session" is still proven to be able to fail (ui-test --fault-inject → 3 reds, exit 1).

Fixed — messages that accused the artifact

• A model file that exists but cannot be parsed was reported as "missing". readJson turned any
  JSON parse error into undefined, so the diagnosis said "项目自己的模型文件缺失 → 补上它" while
  pointing at a file that was right there — following that advice overwrites the user's own file and
  the real cause never surfaces. preload now records {file, why}, and the report has two distinct
  sections: "在，但解析失败（不是缺失，不要照补文件处理）" with the parser's own message, versus
  a genuine missing file.
• Broken lang/*.json was silently skipped. readLang now returns {table, problems}; the
  diagnosis names each unreadable language file with its reason and says what it means (block names
  fall back to raw ids). A missing locale is only reported when no names could be read at all —
  en_us fine and zh_cn absent is a normal state, not a failure.
• atlas.refIcons dropped a cell without saying why. An entry with no png, an empty png, or
  no entry at all now lands in failed with the reason (<id>（没有图标数据：<原因>）); the return
  shape is unchanged.

Added

• texture-path-parity-test.js — the same animated texture (16×N strip + .mcmeta) through
  atlas.scene and atlas.itemIcons must resolve to the same data URL and the same animation facts
  (strip, frames, frametime, order). Reverse fixture: stop the scene path from reading the project's
  .mcmeta (the 0.2.9 bug) and five assertions go red. This is the invariant that keeps the two
  paths from drifting apart again; unifying their implementations is still open.
• model-json-broken-test.js, lang-json-broken-test.js, ref-icons-failed-test.js — each with a
  reverse fixture and a distinct "missing" versus "unreadable" pair of messages.
• The vanilla-parent table has one source. tools/mcart-plugin/vanilla-parents.json is the
  truth: the build generates host/06-vanilla-parents.generated.js (marked generated, ordered before
  its user) and copies the same JSON into the package for the Python extractor, which reads it and
  fails loudly when it is absent instead of falling back to a second table. verify-build
  asserts fragment == regenerated, packaged copy == source byte for byte, and Python's table == JSON;
  --fault-vanilla proves all three can fail.
• The gate list is 50 (was 46); the runner, README and the module-layout doc agree on the number.

Changed — UI (second batch)

Sticky panel header; one control height (26px) across buttons, inputs, selects and icon buttons; the
settings card's switches and hints share an indent column; lists, hotbar, swatches and project
pickers share one thin scrollbar and scroll height; the editor canvas gets breathing room. Colours
remain token-only, with no gradients, animations, glow or stacked shadows.

Known limits

• The icon path and the scene path still have two implementations; the parity gate now makes drift
  fail, but merging them is open.
• The UI is verified by code and tests, not by eye in a browser.

0.3.0：两半源码拆成片段模块；门禁单一入口；产物不再运行时求值

dsh-mc-art-panel 0.3.0

0.3.0

A structural release: the same behaviour, in files you can find things in — plus the build,
gate and dependency hardening that came out of building it.

A correction first

The panel was never broken on DSH 0.2.1. An investigation into "the client UI does not
register" read an empty left sidebar slot (sidebar.panellist) as evidence of a dead client
half; the panel in fact registers in the right sidebar (sidebar.right.pane.tab, key
mc-art.atlas) and in the composer dock (conversation.input.dock, id mc-art-ref), as designed.
The hardening below is worth keeping, but it is not a fix for a missing panel, and this log does not
claim it was.

Changed — the two halves are now fragment modules

• tools/mcart-plugin/host.js (4192 lines) is now 25 fragments under
  tools/mcart-plugin/host/, ordered by host.manifest.json; client.js (4013 lines) is 6
  fragments under tools/mcart-plugin/client/, ordered by client.manifest.json.
• panel/source.mjs is the single concatenation implementation: --write regenerates the barrels,
  --check proves each barrel equals its manifest's concatenation byte for byte, --selftest
  (with a --fault counter-proof) covers ordering, empty-manifest and missing-file errors.
• host.js / client.js remain the complete sources — as generated barrels — so every
  existing gate and test keeps reading the artifact it always read.
• The split is behaviour-preserving and mechanically proven, not judged by sight: batch 1/2
  assert the barrel equals the pre-split source under the one wrapper (installHost, which exists
  so blocks that share closure state can still be moved in order), and the barrel's sha256 was
  asserted unchanged on every later batch.
• What was deliberately not split, with the reason written into the code: apply's closure body
  (client ~3000 lines, and the host handler registry, 736 lines). Fragments share one top-level
  scope; closure bodies cannot cross files without first introducing an explicit panel context
  object, which is a separate refactor — and doing it half-way would trade one tangle for a worse
  interface.

Added — one-command gate runner

• node tools/gates.mjs is now the single source of "the 46 gates": it prints
  绿 / SKIP / 红, lists every red command with its exit code, treats exit 3 as the reserved SKIP
  code, supports --encoding gbk (second pass under the console code page), --only <substr>, and
  --fault, which proves the runner itself can go red. The README's acceptance section now says
  this one command instead of carrying a copy of the list.
• panel/verify-build.mjs gained two assertions that can fail: each barrel equals its manifest
  concatenation (--fault-barrel), and each lib/ artifact equals the regenerated result
  (--fault).
• tools/strip_gate.py and tools/mcart-plugin/report-test.js now compare against the
  concatenated source and cross-check the barrel, instead of assuming a single-file layout.
• tools/mcart-plugin/icon-test.js and icon-render-test.js no longer slice the source by string
  offsets (indexOf('const sub =') … indexOf('const CSS = [')). That made fragment order depend on
  gate internals — reordering files turned them red with ReferenceError: drawItemIcon is not defined. They now evaluate the whole top-level declaration region through one shared helper, and
  throw loudly if the region cannot be found instead of silently evaluating nothing.

Fixed — hardening found while restructuring

• Artifacts no longer evaluate their own source at runtime. Both halves are emitted as inline
  function bodies; verify-build/entry-test assert no runtime evaluation remains, so the artifact
  can be analysed statically and does not depend on a sandbox allowing it.
• Inlining leaked CRLF into the artifacts (about 3,500 carriage returns) — caught by the EOL
  gate the moment the source was no longer wrapped in a JSON string. strip() now normalizes to LF.
• The client plugin declared too little to be safe: it injected only timer and fetched slots
  softly, returning silently when absent — a cold-start ordering change disabled the whole panel with
  nothing on screen or in the log. It now declares inject: ['timer', 'slots'] and reports a missing
  slots loudly.
• dsh.client.inject listed two host-side packages (dsh-api-session-controller,
  dsh-cordis-client-runner), which are not client modules — dependencies that could never resolve.
  The list is now empty, because the real service dependencies are plugin-level ones.
• dsh.compatibility.dsh did not match the running DSH. Under node-semver a prerelease version
  only satisfies a range that names a prerelease of the same major.minor.patch, so
  >=0.1.0-rc.8 <0.3.0 excludes 0.2.1-rc.2. The range now accepts both lines.
• The reference-directory placeholder in settings no longer looks like a Windows user-profile path,
  which the publish-time privacy gate rightly flagged once the source stopped being hidden inside a
  JSON string.

Changed — UI

• Colours now come from host theme tokens (--dsw-alias-* with neutral fallbacks), so the panel
  follows light/dark instead of hard-coding greys; content surfaces (3D viewport, pixel canvas, HUD)
  deliberately keep a fixed neutral dark ground — they are content, not chrome.
• One spacing grid (4/8), collapsed font and radius scales, unified borders, consistent
  hover/active/disabled states, and :focus-visible on every interactive control (custom
  controls previously had no keyboard focus indication at all).
• Empty/loading/error states share one visual language; the HUD's glow was removed. No gradients,
  no animations, no stacked shadows.

Known limits

• Duplicate knowledge is not unified yet: the vanilla-parent table still exists in both
  host/05-model-shape.js and the Python extractor, the icon path and the scene path still resolve
  textures separately, and "cannot read it, return silently" branches remain. Next round starts with
  a drift gate between the two parent tables.
• The UI was reviewed by code and tokens, not by eye in a browser: the theme-token inspect provider
  timed out, so token names were read from the shipped DSH artifacts.

0.2.9：项目自己的动图，面板从来没当成动图；而且那句提示在指责文件

dsh-mc-art-panel 0.2.9

0.2.9

Fixed

• An animated texture in your own pack was never animated in the panel, and the panel blamed the
  file for it. animationsFor() consulted only the reference textures' animation table, and
  preload() collected the project's textures without ever looking for <name>.png.mcmeta — the
  word mcmeta did not appear in the host at all. A 16×192 strip (12 frames of 16, frametime 4)
  was therefore drawn as one 16×192 texture, which is the blur you saw, while the game read the
  .mcmeta and animated it correctly. The host now reads the project's .mcmeta (its own IHDR size
  reader; the same acceptance rules as the extractor's animation_of) and animationsFor() consults
  both tables.
• The message no longer accuses the file of something the tool never checked. "shaped like a
  strip but has no animation description" was printed whenever the shape looked like a strip and no
  animation entry existed — including when the .mcmeta was right there. There are now three cases
  saying three different things: a valid .mcmeta → nothing is shown; .mcmeta present but
  unreadable (bad JSON, no animation section, or the image is not a strip) → "this texture
  has an animation description file, but I cannot read it: <path>.mcmeta — <the real reason>"; genuinely absent → the shape-based message, which now also names the file to add.
• Two unrelated drifts surfaced by gating this: check_jdk_test.py had been rewritten to capture
  output without declaring how to decode it, so encoding-test.js --fault could not catch its own
  reverse fixture (the static rule now requires encoding=/text= whenever output is captured);
  and the host's line endings had been normalised to CRLF, which made multi-line --fault patches
  match nothing silently (readHostSource() now normalises to LF). Both guards reported
  "did not take effect" rather than turning green.

Added

• tools/mcart-plugin/project-animation-test.js: fixture A (valid .mcmeta — must be 12 frames and
  must produce neither message), B (absent — may produce the shape message), C (present but broken —
  must produce the truthful message and must never say "no animation description"), D (a plain
  16×16 — silent). --fault (remove the project .mcmeta reading) turns 8 assertions red.
• panel/ui-test.mjs gained the same three cases plus --fault-animnotes, whose reverse fixture
  turns exactly case C red while case B stays green — the fallback message is correct when the
  host has nothing to say, so only "the host spoke and the client ignored it" may fail.

Note

The .mcmeta path was already parsed on the reference side (that is where the jar's animations come
from), which is why this only broke for textures in your own pack — precisely the case the panel
exists for.

0.2.8：引擎写出的方块模型在游戏里不能用，而所有门禁都说没问题

dsh-mc-art-panel 0.2.8

0.2.8

Fixed

• The block models the engine wrote were unusable in the game, and every gate said they were
  fine. write_block_model passed the spec's elements through verbatim, so a face written as
  "stone" — a key into the model's own textures map, which is the spec's convention and what
  the engine's own audit validates — stayed "stone" in the emitted Minecraft model, where a value
  without # is a path: <namespace>:textures/stone.png, which does not exist. Every face lost
  its texture, so the model was invisible in game and undrawable in the panel. Measured on a real
  project: 42 references in one hand-authored model, 38 in another — and those two were the
  only broken ones, because they were the only ones going through the block-entity path.
  Face textures are now emitted as #key; values that are already #… or namespaced paths are left
  alone (vanilla writes those too).
• The artifact is now validated by reading it back. validate_written_model re-reads the model
  it just wrote and requires every face to be a #variable present in that model's own textures,
  or a path the pack really ships; otherwise it raises, so an illegal model cannot leave the
  generator. It sits on the generation path, ahead of the preview — because the preview reads PNGs
  from disk and therefore renders correctly even when the written model is broken. That split is
  exactly why this shipped: the picture was right and the artifact was wrong.
• The engine's own example examples/example_block_entity/desk.json is a spec whose generated
  model was broken; it now emits #wood/#paper, and the emitted model is read back and validated
  (20 faces, zero bare words).
• A second source of truth, removed by choice. A spec may now declare a texture's real asset id
  ({"path": …, "id": "ns:block/real_name"}), in which case nothing is copied. Copying stays the
  default, but when the spec key differs from the source file's name the copy is reported with
  its consequence: the pack ships those bytes renamed, so anything referring to the real asset id
  will not find them.

Added

• The panel now names this mistake instead of sending you hunting for a missing PNG. A face
  texture that is a bare word and matches a key in the model's textures map is reported as
  texture-missing-hash: it names the face, says the bare word is being read as a path, lists
  the model's texture keys, and says "you probably meant #stone" — adding that the block is
  broken in the game too, not only in the panel. A #variable that is not in the map is a separate
  reason (texture-variable-unresolved), and a genuinely missing file stays textures-unresolved;
  the three no longer share one sentence. tools/mcart-plugin/texture-syntax-test.js covers all
  three plus two legal forms ("#stone", and a bare path like block/stone_tex) so the advice
  cannot fire on correct input.
• Unresolved textures that are the project's own no longer report no-reference-directory; a
  missing PNG in your own pack has nothing to do with whether a reference directory is set.

Changed

• Vendored mc-art snapshot: 49793ad → a6e224c.

Note

mc-art doctor caught the installer's clone being one commit behind before this release — the
commit being this fix — and said which side was stale. Fast-forwarding that clone is now part of the
re-vendor step; without it the package would have shipped the broken model writer again.

0.2.7：母模型被取成空壳时整条链就断，而报告说「原因不明」

dsh-mc-art-panel 0.2.7

0.2.7

Fixed

• A block whose model parents minecraft:block/cube_all could report "the chain is complete, but
  not one face was drawn" with 原因：unknown. Two separate defects sat behind that one message:
  • The parent fallback chain accepted any truthy result as "resolved", so a parent fetched as
    an empty shell ({} / {textures:{}}) short-circuited the built-in vanilla parent table
    and the chain ended with no geometry. An empty shell is not a resolution. This is not a
    hypothetical shape: the game's own block/block carries only display/gui_light — no
    elements, no parent — so an ordinary chain really does contain one.
  • no-quads was only set when elements !== undefined, so "the whole chain has no elements" fell
    through to unknown — and unknown is not a diagnosis. It is now no-geometry, and the
    text names the real cause (a parent that could not be fetched) and says explicitly that this is
    not evidence of a block entity.
• The report's line about the built-in parent table printed "（这次没走到）" whenever nothing was
  missing — i.e. it claimed the table had not been used even when the table was what resolved the
  chain. It now says what actually happened: the keys it hit and which one caught the chain, or that
  it was consulted and did not have the key, or that it was not needed at all.
• tools/mcart-plugin/parent-fallback-test.js: a self-contained project whose pack contains an empty
  shell must still draw; the control (no cube_all in the pack, so the built-in table must do the
  work) must not go red; and the reason must not be unknown. The reverse fixture turns three
  assertions red.

Note

The build that produced the report predated vanillaBare() (present from 0.2.3): a project model
written the way vanilla writes it resolved to nothing there. Worth recording because the panel's
host half only reloads when DSH is fully restarted — approving a new version is not enough on
its own.

Changed

• Vendored mc-art snapshot unchanged (49793ad): this release is panel-side only.

0.2.6：那条嵌入带量错了东西；判定器在含空格路径上跑不起来

dsh-mc-art-panel 0.2.6

0.2.6

This release is mostly one thing a user's own second session measured and wrote down: a
757-line log of what actually cost time while building a mod with these tools. Its first
finding was a defect in a gate we shipped two releases earlier.

Fixed

• The embedding band was measuring the wrong thing. accent_base_gap's default 6..24 came
  from iron_ore, and it was the mean of a whole ore cluster, not of a speck. Measured across
  all eight vanilla ores the figure runs −75.7 (redstone) … +106.6 (gold) and none of them
  falls inside 6..24. Using the default as an ore's target compresses a deposit that should read
  as bright into a recessive smudge — the session did that for seven rounds. The constants are now
  named for their owner (ITEM_ACCENT_GAP*), the vanilla ore range is published as a fact,
  accent_base_gap_min/max default to None so the engine can tell a default from a
  declaration, and an asset declared as a deposit that declares no band is refused with the
  way out in the message. The default was not loosened: that would refuse items.
• mc-art gap-from-refs measures the band from your own references, and reports the
  per-deposit figure (the one to declare) separately from the per-pixel boundary figure —
  stating that averaging a cluster first is exactly how 12.24 was produced, and printing its mask
  rule, because a different mask gives a different number.
• The judge could not run on a path with spaces. list2cmdline() was handed to subprocess as
  a single argv element, so cmd received a literal \"…\" and answered
  '…' is not recognized — which looks like a Gradle failure. The user's own project path
  (Release 2.8.3) has a space, so this was blocking. Windows now gets one command line.
• --fault returned 0 even when it caught nothing, which made "the checker can say no"
  unfalsifiable. It now returns 4 when the injected fault goes undetected (1 stays "the verdict
  did not pass"), and says that the tool's own exit code carries no verdict.
• --fault's injection point was hardcoded to the example mod's expression, so any new project
  had to write ExampleMod.EXAMPLE_BLOCK verbatim — polluting the subject to satisfy the tool.
  --fault-find / --fault-replace were added; the old pair stays as a compatibility default and
  the failure now points at the flag.
• The judge did not manage Gradle's home, and the default one is read-only to a sandboxed
  process, so the wrapper died creating a lock file. It now probes candidates, really creates a
  lock file to prove writability, prints which home it chose and why each other was rejected, and
  passes GRADLE_USER_HOME. (gradlew's -g cannot fix this: the wrapper parses it only when it
  precedes the task name.)
• geometry.uv_regions silently cancelled contour conformance (appearance_only plus declared
  regions degrade an item to a full opaque quad — measured 0 transparent pixels where vanilla has
  188). The two declarations cancelling out are now reported.

Added

• mc-art doctor — platform, engine root, module count, which interpreter works and why each
  candidate was rejected, whether the three entry points really run, line endings against
  git check-attr, writability, and whether the engine copies on this machine agree. On its first
  run it caught a stale ~/.dsh/skills/mc-art and said which side was behind.
• scripts/check-ore-gap.py (an undeclared ore is refused, a declared one passes; both
  directions proven) and scripts/check-tiling.py — a real tiling-seam gate counting pattern
  pixels on row 0/15 or column 0/15 against a declared tiling_min_margin. Six of eight vanilla
  ores keep a margin of 1; only emerald touches the edge, and the message says so.
• tests/test_doc_counts.py — the module count a document claims is compared against
  mc_art/*.py, so the number can no longer drift (it caught doctor.py immediately: 34).

Changed

• The skill now teaches the path that was missing. Measure the reference corpus first —
  colour and shape grammar (per-face blush count, cluster size, fill ratio, thickness, luma
  levels; the vanilla ore contract measured across seven ores is 6–9 clusters, 4–29 px,
  fill 0.70–0.84, thickness 5–9, 2–5 levels) — then state per-asset-class success criteria, then
  draw. Plus: which gates are declarable versus fixed, that "attached a reference" is not "used the
  reference", that layer matching runs both ways, and that the band a category needs comes from
  its own references. windows.md gained the measured PowerShell traps (write a .py, never
  inline; Set-Content -Encoding UTF8 writes a BOM that breaks json.load), and traps.md gained
  the one that quietly destroys good pixels: a "repair" step written before its premise was checked.
• Vendored mc-art snapshot: 427fd01 → 49793ad.

0.2.5：你自己包里的贴图保存不了（那条消息在撒谎）

dsh-mc-art-panel 0.2.5

0.2.5

Fixed

• Every texture in your own project pack refused to save. The panel said "this texture handle
  is not a relative path inside the project pack" while the path plainly was inside your pack:
  the sentence was true about the shape and false about the reason. The host stored your
  project's own textures as absolute paths (preload()), that value travelled to the editor as
  its save handle, and the client's safety check rejects drive-letter paths — so no texture in your
  own pack could be saved, and the write path was never reached at all. The host now stores
  pack-relative handles and resolves them against the project root when reading; the in-pack test
  normalises separators, case and ./.. and compares real paths, and it accepts an in-pack
  absolute path as well, so a stray absolute handle cannot lock you out again. The rejection
  message now names what is actually wrong — empty / NUL / drive-letter / leading slash / .. /
  outside the pack, five distinct reasons — instead of one sentence that blames your own pack.
• tools/mcart-plugin/texture-edit-test.js: the handle must pass the real client check
  (safeTextureHandle, pulled out of client.js rather than re-implemented) and the host must
  actually land the bytes in the project pack (before/after compared). The reverse fixture puts the
  absolute path back and turns six assertions red — including the message users actually saw,
  verbatim — with the file unchanged at 224 B, i.e. not one byte written.

Changed

• Vendored mc-art snapshot unchanged (cf39eba): this release is panel-side only.