Purenista MTrade Reference Guide

Fixed: the artwork check told you to enable a feature that no longer exists

• When the check decided there was no reliable evidence of text, the explanation ended with “pick ‘text / layout’ under (2), or turn on AI vision assist”. The AI vision assist interface was removed entirely in v5.2.4 — there is no switch to turn on and nothing to configure. It asked people to do something impossible, and left them thinking they had missed a setup step.
• That sentence is gone, leaving only the thing that actually works. The dead code went with it: the branch that would have said “text located by AI vision assist” and could never be reached, five strings that were never displayed, and the empty entry in the report's tab map.
• An old comment also survived in the source describing how images were sent to an external endpoint for semantic analysis. The feature was long gone but the comment read as though it were still there — exactly the impression v5.2.4 set out to remove. Replaced with one line: an uploaded image never leaves the browser.

Tidied: the Cloudflare files are no longer bundled

• Site-wide ranking needs no setup, so those two files only made it look like there was homework outstanding. Removed; ask if the day ever comes. The site itself is unaffected — everything it reaches out to (counters, fonts, the anime face model) is automatic, and none of it requires you to sign up for or host anything.

Build tag updated to v5.4.5. ZH and EN in sync.

Fixed: site-wide ranking still would not load — a different approach rather than another guess

• The last version fixed "nobody has clicked yet" being read as a failure, and it still would not load on the live site. That rules out the 404 case and means the bulk-read URL itself does not work — either it does not allow cross-origin requests, or the legacy API dropped it. There is no way to tell which from outside.
• So it no longer depends on it. Every time the site adds one to a tab's counter the reply already carries that tab's site-wide total, and that call is known to work — the view counter has used it for a long time. Each reply is now remembered, and when the bulk read is unavailable the ranking is built from those real site-wide figures instead.
• The cost is stated plainly on the card: in that situation only tabs you have opened can be listed, though every figure is a site-wide total rather than a local one. The more you visit, the more complete the list. If the bulk read ever starts working it goes back to the full table automatically.
• When nothing can be read at all, the message now includes the actual error instead of a vague "try again later" — so the next time something breaks it is visible rather than guessed at.
• A circuit breaker was added too: after two consecutive bulk-read failures it stops retrying for the rest of the visit, instead of firing dozens of doomed requests on every switch.

Build tag updated to v5.4.4. ZH and EN in sync. After uploading, open a few tabs before checking the site-wide ranking — a figure only exists once someone has clicked.

Fixed: the site-wide ranking always said it could not be read

• On a freshly deployed site the site-wide ranking never loaded. The cause was “nobody has clicked this yet” being read as “the request failed”: the counter service returns 404 for a counter that has never been incremented, because it does not exist yet. The code treated any non-200 as an error, so on a site with no clicks all 22 tabs 404'd and it looked like the whole service was down.
• A 404 is a zero, not a fault. Only a network error or an unexpected status counts as a failure now, and with no data the panel says the site-wide table is still filling up rather than telling you to try again later.
• The irony is that the comment in the original code already said “a 404 is a zero, one missing tab must not sink the whole table” — the comment described the intent and the code did not implement it. Re-verified against four scenarios: nothing counted yet, partial data, fully offline, and the read URL changing shape.
• One safeguard added while there: the documentation writes the read endpoint both with and without a trailing slash, so a failure on the first form now retries the second automatically instead of waiting for a release.

Fixed: the card's description still described the old setup

• It still read “connect a free cloud counter to switch to site-wide”, but site-wide no longer needs anything connected. It now just explains the difference between the two modes.

Build tag updated to v5.4.3. ZH and EN in sync. Site-wide numbers only appear once someone has actually opened a tab — the first one counted will be yours.

Site-wide ranking works out of the box now — nothing to set up

• It used to require creating a Cloudflare Worker, binding a database and pasting the URL back into the source before it did anything at all. It is already counting by default, using the same counterapi service this site already relies on for the view counter, the ratio tool and the colour generator — same namespace, no account, no key, no deployment. Switch to the site-wide tab and it is there.
• The limits, for the record: counterapi allows 30 requests a minute, counted separately per URL path. Twenty-two tabs are twenty-two paths, and reading the whole table costs one request against each, so there is plenty of headroom.
• The Cloudflare Worker has not been removed, it is now an optional upgrade: fill in the URL and it takes over, turning 22 reads into 1 and putting the data somewhere you control. Worth it only if the site ever gets busy enough for counterapi's limit to matter; ask for the code if that day comes.

Fixed: the first site-wide render fired twice the requests it needed

• Both the render path and the boot path asked for the full table, and the cache was still empty when the first one asked, so both genuinely went out — 44 requests instead of 22 on counterapi. Only one read now runs at a time; later callers join the one already in flight.

Also fixed: the optional Worker lost hits as soon as more than one person used it

• The Worker kept all 22 tab counts in a single KV key and incremented by reading the whole object, adding one and writing it back. That tests perfectly on your own and breaks with company: KV has no atomic increment, so two people clicking at the same moment both read 100, both write 101, and one hit vanishes. The more popular the site, the more it undercounts.
• Two more limits that fail silently: one write per second per key (all 22 tabs shared one), and 1,000 writes a day on the free plan — after which counting simply stops for the day with no error, the numbers just stop moving.
• Now on Cloudflare D1 (free SQLite): one atomic SQL statement, 100,000 writes a day. Setup gains two steps — create a database, run one CREATE TABLE — and in exchange the numbers are actually right.
• The Worker now has an ORIGINS allow-list at the top; fill in your own domain and it only accepts browser requests from your site. Honestly: it does not stop anyone calling the API directly. Real protection needs a captcha or accounts, which is beyond a widget like this.

Fixed: three front-end problems in the ranking list

• In site-wide mode the entry names come back over the network. An unrecognised key used to be printed as-is, which handed a remote server a way to write markup into the page. Only tabs the site actually has are accepted now, everything else is dropped — which also handles counts left in the database for tabs that no longer exist.
• Ties were sorted with a hard-coded Chinese collation, on the English page too. It follows the page language now.
• The 60-second anti-spam timer lived only in sessionStorage, inside a try that swallowed the failure — so in private mode, or wherever storage is blocked, there was no throttle at all and every tab switch posted. Memory holds the floor now; storage only carries it across reloads.

Restored: last round's art-check fixes had not been carried over

• Four fixes derived from cross-checking twenty real pieces were lost in the v5.4.0 merge and are back: (1) the score and the fix list now agree on what is not being judged (six of the twenty were wrongly called out of focus, the lowest measuring 63); (2) the alignment check reads the glyph detector's own output instead of the block test that treats lace as text (one piece with no lettering measured 52% coverage), with tightened thresholds that take false reports from nearly-always down to one; (3) collages and multi-panel sheets are detected automatically (fired on the one 2×2 sheet and nothing else); (4) a mount on one side only can be recognised, at a raised threshold of 150.

Build tag updated to v5.4.2. ZH and EN in sync. The site-wide ranking needs no configuration at all — uploading the files is enough to start it counting.

Feature ranking: Mine / Site-wide

• Toggle between local and site-wide; local remains the default.
• Connect a free Cloudflare Worker to aggregate anonymous tab opens.
• Nothing is sent while GLOBAL_RANK_ENDPOINT is empty.
• See cloudflare-rank-worker.js and the setup notes in the zip.

Build tag v5.4.1.

Art check: full judgement overhaul (merged into later base)

• Technical / style split: no hard faults → “composition holds, no mandatory fixes”.
• Illustration / design safety net: only severe alignment or extreme clutter stay technical.
• Near-mono / high-key / cinematic / Morandi clear hard list; focus, space, UV, subject scale → style notes.
• Diptych, collage, circular, banner isolation, magazine type, line art no longer over-penalised.

Build tag v5.4.0 (later content base + judgement merge).

Expanded: the palette and text libraries behind the random generator

• Palettes 45→70, themes 58→88, character traits 70→100. The new palettes lean towards tones the game's wardrobe actually has: Oat & Cream, Misty Pine, Moon White, Teahouse Lamp, Blue Hour, Paper Craft, Sea Salt Air, Stone Garden, Misty Lilac, Library Oak, Graphite Sketch, Mushroom Forest, Ash Rose, Aurora Night and more.
• The 30 new traits keep to concrete behaviour rather than personality labels (paces without noticing while on the phone; always leaves the umbrella at someone else's place; says “it's fine” when it very much is not; gets more polite the angrier they are) — far more useful for an OC than an adjective.
• Every new palette carries ten keywords, so they are reachable by typing a keyword and not only by rolling.

The whole weight table was recomputed, not just extended

• How often a palette comes up is a property of the set, not of any one entry. The 25 additions are themselves mostly muted, so weighting them by the existing rule would have quietly moved the muted share from 58% to 63% — nothing breaks, the rolls just slowly stop varying. After recomputing, the library sits at 58% muted / 41% mid / 1% vivid, exactly where it was before.
• This is the lesson written down when 20 palettes were added in v5.3.0, applied this time.

Fixed: keyword matching was won by whichever palette was declared first

• Matching scored one point per matched term and broke ties by declaration order. But keywords are shared on purpose — eleven palettes list “gentle” — so a broad word tied across a dozen palettes and always returned the same one. Going from 45 to 70 palettes would only have made that worse.
• Two changes: a longer matched term scores higher (a longer term is the more specific claim, capped so one long phrase cannot outvote several real hits), and a genuine tie is now broken at random instead of by declaration order. “Stone garden”, “second-hand bookshop” and “cyberpunk neon” now land reliably on the right palette, while “gentle” and “forest” spread across sensible candidates.
• Same bug the title generator had in v5.2.9 — this time on the colour side.

Fixed: a large block of data had never been translated

• The v5.3.0 expansion was added in Chinese only: 28 palette names, their five colour names and ten keywords each, plus 28 themes and 40 character traits, were all still Chinese on this page. All translated now — keywords included, since without those an English search could never reach those palettes at all.
• One more oddity: the keyword list for the pearl palette contained a literal string reading typo, evidently left behind while editing. Replaced with a real word.

Build tag updated to v5.4.0. ZH and EN in sync; all 88 themes point at palettes that exist, all 70 palettes have five swatches and a weight, and no theme or trait is duplicated.

Improved: the mobile tab menu is now grouped and collapsible

• Twenty-two destinations in one flat list meant a long scroll — collapsing the nav bar was supposed to make choosing faster, and it did not. The sheet now shows the seven groups (Guide / Trading / Spending / Reference / Create / Learn / About) and slides exactly one open at a time. Opening the sheet expands whichever group holds the tab you are on, so the common case is still one tap; each header also shows where you are and how many items it holds.
• The group headings used to be unclickable decorative text. Now that they are real buttons, the panel's accessibility role moved from listbox to menu, with aria-expanded and aria-controls added.

Fixed: two parts of the artwork check had never actually run

• Text alignment check: the result was read about ninety lines before it was computed. JavaScript's variable hoisting means that is not an error — it just quietly reads “undefined”. So the “these text blocks share no alignment grid” advice had never once appeared, and neither had the rule that caps the composition score at 65 because of it. Moving the whole text-detection pass ahead of the scoring fixed it.
• “Add an accent colour” advice: the condition compared the colour-richness number against the translated text label. Those are never equal, so that suggestion had never fired either.
• Both are the same kind of bug: nothing breaks, it just silently does nothing — which is why they sat there so long.

Fixed: illustration mode said it wasn't judging, then marked you down anyway

• The report is explicit about it: a large flat white is paint, not a blown highlight, and flat shading has no focus to miss — so illustration mode declines to judge them. The score never followed. Worse, illustration mode rewrites the “blown / crushed” verdicts into the neutral “high key / low key” wording, and neither of those strings matched any scoring branch, so both dropped straight to the failing value. A deliberate white-on-white piece was excused in the prose and marked to the floor in the number. Criteria that do not apply now cost nothing; criteria that genuinely reached a verdict still count.
• This lifts illustration scores somewhat — they were being held down by penalties that should never have applied.

Fixed: things you could tap that went wrong

• Kaomoji were shattered: the splitting pattern also cut at the spaces inside a face, so ( ; ᴗ ; ) became five separate buttons — “(”, “;”, “ᴗ”, “;”, “)” — and copying gave you a fragment. Symbols are separated by one space and kaomoji by two; each kind now uses its own separator.
• Flower buttons were destroyed on first tap: the “✓” flash overwrote the button's text content, wiping out the colour swatch and layout inside it, and a second later only plain text came back. It now snapshots the full contents and restores those.
• The colour-inspiration keyword wasn't escaped before going into the page. Everywhere else on the site handles this; this one spot was missed.
• A failed image gave no feedback at all: a corrupt or unsupported file left the page silently unchanged, so you could not tell the upload had failed. It now says so, and releases the temporary URL it was holding.
• Colour richness was inflated when no subject was found: the colour-gamut span multiplied two negative sentinels into a large positive, so “almost no colour information” read as “very rich colour”, and the style verdict followed it.
• The random picker could retry forever in one edge case (all candidates identical); it now gives up after a bounded number of tries. Two elements that had never had any styling also got some.

English version: the missing translations

• Symbols, kaomoji and flower meanings were untranslated — the category names and all 32 flower names and meanings were still in Chinese while the surrounding copy was English. All translated; the Japanese hanakotoba marker changed from 【和】 to [JP], and the copy format now uses English punctuation.
• Chinese lookup keys were reaching the screen: the motif and mood tables are keyed in Chinese in both builds because the matcher indexes those keys — but a key is not a label, and the seasonal chips, the generated hashtags and every creative-workspace option came out in Chinese. What is stored is still the key; what is shown now goes through the language table.
• A few more mixed-language leftovers: “非Symmetrical”, four fully Chinese captions inside the typography diagrams, one English sentence with a stray Chinese character in the middle, and a Korean character mistyped into one of the motif entries.

Build tag updated to v5.3.3. ZH and EN in sync.

New: smart guide and on-device feature ranking (client-only)

• Smart guide: as its own first tab “Smart guide”; pick a role (newbie / gacha / trade / brand / create) for a suggested tab path with one-tap jumps. Runs entirely in the browser — no backend, no login. Last role is remembered locally.
• Feature popularity (this device): top tabs by how often you open them, stored in localStorage only — never uploaded. Stats can be cleared in one tap.
• Both features are deliberately server-free: the guide is a static path table; the ranking is a local counter. No account, no tracking.

Build tag v5.3.2. ZH and EN in sync.

Fixed: art-check false positives (cross-checked on real plates)

• Warm / cool label: low-chroma grey and silver illustrations were often tagged "warm" because only high-saturation hues voted, so a few skin or accent pixels pushed the ratio over the line while cool greys abstained. Near-neutral pixels now vote via the Lab blue–yellow axis, low-chroma frames lean on measured Lab, and the "warm" DNA line requires both ratio and Lab to agree.
• Unity and variety: restrained mono, cinematic and Morandi plates score low on variety by design; the geometric mean then looked like a fault. Low-chroma / high-harmony work no longer lists "unity and variety not differentiated" as a defect — that is the style, not an error.
• Scattered focus: soft light, full moons and halos spread brightness so area-holding-half-the-weight looked "unfocused". When a face is detected, or a centre / radial read is strong, the hard "no landing point" penalty is skipped.
• Uneven negative space: branch frames, dark mass and magazine title bands are often deliberate asymmetry. In design mode with text, vertical gaps are no longer a hard fault; in illustration mode with a centre / radial pattern, mild left–right imbalance is heavily down-weighted.
• Black screen / visionState: a leftover undeclared variable could blank the result panel; it is now initialised, and errors draw text instead of a black canvas.
• Aesthetics sub-tabs unclickable: listeners bound before lazy content injection never attached. Click handling is now document-level delegation so sub-tabs work after hydrate.

Build tag updated to v5.3.2. If the page still looks old after deploy, confirm GitHub Pages has the new files and force-refresh / clear cache on mobile.

Fixed: "Copied" appeared even when the copy had failed

• Following the data-c collision fixed last version, all three copy features were audited. The character generator was correct; the title generator and the workspace plan card were not - errors swallowed by an empty catch, the confirmation sitting outside then() so it showed regardless of the result, and no fallback at all.
• navigator.clipboard only exists in a secure context, so anyone opening the file straight off disk lost copying entirely while the page kept reporting success. All three now share one helper: use the API where present, fall back to execCommand otherwise (including the setSelectionRange iOS needs), and say "copy failed" plainly when both routes fail. The confirmation only appears once the write has resolved.
• Also fixed: clicking a copy button twice quickly recorded "Copied" as the button's original label, so the text never went back.

Expanded: every random pool

• Palettes 17→45, themes 30→58, character traits 30→70, motifs 20→80, moods 8→28, title templates 47→519, backgrounds 10→32, accessory types 10→30, framings 7→19.
• A player reported the palette dice always landing on vivid schemes while the game's wardrobe is mostly muted. Measured: only one of seventeen palettes averaged under 30% saturation, and every palette was equally likely. After adding muted palettes and weighting the draw, muted rose from 6% of rolls to 50% and vivid fell from 18% to 2%.
• When twenty more were added later the weight table was recomputed in full - the easy mistake at that scale is letting the muted proportion quietly drift back.
• The forty new traits describe behaviour rather than personality labels, which is more use when writing an OC.

Build verification added: all 58 themes point at palettes that exist - a wrong reference would not break the page, it would silently show no swatches. Eleven clipboard tests cover insecure contexts, denied permission, both fallbacks failing, and double clicks.

Fixed

• Typing "beach" returned "sugar". Not a mis-read - the word was not in the table, and an unrecognised keyword fell through to a random pick. Of twenty ordinary words tested, eighteen were unrecognised. The real problem was not pool size but silently substituting something else, so an unknown keyword is now used literally.
• Results felt repetitive for two reasons: all six titles shared one motif, and no record was kept of templates already used. Now the motif varies line by line and the last thirty templates are skipped - eight consecutive rolls produced 48 titles in 47 distinct shapes.
• Tags padded themselves with an unrelated motif; searching "beach" produced #Piano. They now follow the keyword.
• Also fixed a problem that only appeared once motifs grew from 20 to 80: matching scanned the table in order and took the first hit, so the single character 光 intercepted 月光. Longest match wins now. The bug was intermittent - two failures in five runs.

Added

• Decision helper (at the top of the workspace): four either/or questions - figure or scene, light or dark, cute or grown-up, warm or cool - fill the seven workspace fields in about ten seconds. Each of the sixteen combinations maps to a hand-written direction rather than four independent dice, so results never collide the way "cute x ruins x cool" would. Every one carries a note specific to that combination.
• Workspace: subject, mood, palette, background, framing, accessories, title. No fixed order; four filled fields produce a copyable plan card. Options are given by type, never by in-game item name - items go stale with each update, categories do not.
• Naming and tags: five registers, with tags drawn from the keyword and the current month. No "trending tags" - that needs a server watching everyone's posts, and an unmaintained "trending" list is worse than none.
• Workflow checklist: ten items across before / during / before posting. Each maps to something the artwork check actually measures, so the advice and the later verdict are one standard stated twice.

Build verification added: every feature declares both its markup and the code that drives it, and the build fails if either is missing. Twice a feature's JavaScript had been deleted while its interface remained - the page looked fine and the buttons simply did nothing.

Performance

• The aesthetics column templates (246 KB) plus the trivia, brand-data and feedback tabs (129 KB) moved to an external file, loaded the first time one of those tabs is opened. First load 783 KB to 414 KB, transferred (gzip) 250 KB to 143 KB, DOM elements 2,823 to 1,460.
• Loaded with a <script> tag rather than fetch, because fetch is blocked under file:// - opening index.html straight off disk would otherwise show a broken column.

Fixed

• Three diagrams in the aesthetics column drew their arrowheads in the wrong colour: each defined <marker id="a2">, but url(#a2) resolves document-wide to the first one. All 275 SVGs now reference only their own definitions.
• Removed 13.5 KB of dead code: the AI vision assist interface had been taken out but its implementation remained, including api.anthropic.com - which reasonably suggested the site sends artwork to a third party. It does not, and now it does not look like it might.

Fixed

• Found why the Trivia page's "Common" sub-tab (Naming & language / Allusions / Stories) wouldn't open and popped up "Copied" instead: those buttons used the attribute name data-c, which collided with the same attribute name used by the "tap to copy" feature in the Naming & Tags tool — a page-wide copy listener was grabbing them too, regardless of context. Quick Calc's tab buttons use the same attribute name and were being caught the same way.
• Two-part fix: ① the copy listener now only fires inside the Naming & Tags tool's own container, not anywhere else on the page; ② the Trivia "Common" sub-tab was renamed to data-v, matching its sibling tabs, removing the name collision at the source.

Both language versions (including content-zh.js / content-en.js) were updated and syntax-checked.

Fixed

• The brand column in the BRAND basket forecast table is back to full brand names (Deux #16, melem #5…) instead of short codes. The previous fix squeezed columns into fixed percentage widths, which is what forced the short codes in the first place; this version gives the table its own minimum width (800px) so on narrow screens the whole table scrolls left/right instead of sacrificing readability. The code legend above the table is now a swipe hint instead.

Tested at a 390px-wide mobile viewport — full brand names no longer overlap, and the table scrolls to show everything.

Fixed

• The brand column in the BRAND basket forecast table was still overlapping with the date column on narrow screens ("Deux #17" running into the date's "04" and reading like "07/04"). Brand labels are now short codes — D = Deux, M = melem, by = By(e), the number is the basket count (e.g. D17, M5, by4) — and the column width dropped from 14% to 9%, with the extra space going to the date and confidence columns. Added a one-line legend above the table.

Tested at a 390px-wide mobile viewport — no more overlap.

Changed

• Removed the "AI deep review" and questionnaire-scoring logic from Character check — it required players to bring their own Anthropic API key, which almost nobody has, so the core metrics (memorability, development potential) could only ever return low-confidence local guesses. Just cut it rather than keep a dead-end feature.
• Renamed the tool to Random Character Generator. The old 10-field questionnaire is now a rollable character sheet: fill blanks only (leaves what you wrote untouched), reroll everything (overwrites the whole set, new), copy character sheet (one-tap copy to clipboard, new). Random generation now covers all 10 fields instead of 6 — added dedicated pools for name, backstory, relationships and hook.

Both language versions updated together and tested end-to-end (generate / reroll / copy) in a real browser with no errors.

Changed

• Random Inspiration used to bundle theme / trait / composition / palette into one draw — if you didn't like one part, you had to reroll the whole set. It's now four independent dice: keep the ones you like and only reroll what you don't, with a "Reroll all" button still available for a full redraw.
• Palette is no longer tied to the theme — it rolls independently now. Also added 10 more entries each to the theme and trait pools, so combinations feel less repetitive overall.

Fixed

• The Color Inspiration generator used to show its 5 colors as equal-size swatches, which read like a flat 5-color mix and didn't reflect the basic "more dominant, less accent" color principle. It's now a single bar split by proportion (46/26/14/9/5%), with each color's role labeled below (Dominant / Secondary / Tertiary / Accent × 2) — closer to how a real palette is actually used, and better-looking too.

Fixed

• The "Brand" column in the BRAND basket forecast table was being forced to not-wrap by a site-wide numeric style (td.num), so longer names like "Deux #18 ex" overflowed and overlapped the "Predicted window" column next to it. Brand column now wraps normally again — no more overlapping text.
• "Trade Smarts" → "How to decline politely" card: reworded two of the example phrases to sound more direct and closer to how the situation actually plays out.

Fixed

• The "Confidence" column in the BRAND basket forecast table was too narrow — short values like "Confirmed" could get squeezed into a single-character-per-line stack. Widened the column (12%→20%) and set data cells to not wrap, so short confidence values now stay on one line.
• The other three columns (brand / window / theme) were adjusted slightly to compensate; the theme column still gets the most space.

Fixed

• Wrapped the "BRAND basket forecast" table in a horizontal-scroll container: even if the column-width math falls short in some edge case, the table now scrolls sideways within its own box at most, instead of stretching the whole page and pushing the "Confidence" column off-screen.

If the page still looks like the old layout after this update, double-check that the file live on GitHub Pages is this version, and hard-refresh / clear cache on your phone's browser — GitHub Pages can take a few minutes to roll out changes.

Fixed

• Quick Calc → Ratio calculator: item type / condition / what-the-other-side-offers are now native dropdowns instead of button groups — no more tapping through a row of pills.
• Schedule → BRAND basket forecast table now has fixed column widths (via colgroup), so the "predicted theme · colors" column no longer gets squeezed into a one-character-wide wrap that pushed the whole table off-screen; the date column is narrower and wraps to two lines, freeing up space for the column with the most text.
• On narrow screens (≤680px), the row of 17 tab buttons at the top is replaced with a single "jump to section" dropdown, removing the need to scroll sideways through it; wider screens keep the original pill row. Both share the same switching logic and stay in sync.

All three came from the same feedback: swipe-based selectors felt tiring on mobile, so this pass moves them toward native dropdowns. Both language versions were updated and syntax-checked.

Added

• New tab "Color Inspiration" (right after Character Check): merges the palette generator and the random-inspiration draw into one tool with four entry points, each returning a matching style palette — ①Keyword input, matched against the closest style palette; ②Extend from an uploaded image, pulling its dominant colors to find the nearest style; ③Paste a character description to reverse-match a style, with a one-click "pull in my Character Check answers" button; ④🎲Random inspiration, drawing a theme + character trait + composition suggestion + palette all at once — the composition suggestion reuses the same 18 composition types already documented in the Aesthetics column.
• The palette library covers 17 style moods (ethereal, dark fantasy, cyberpunk, candy, forest, ocean, steampunk, gothic, wasteland, imperial court, neon nightlife, desert, wafu, mecha, polar, circus, retro), 5 colors each; the theme pool and character-trait pool each hold 20 entries; keyword matching ships with a synonym dictionary covering both Chinese and English input.
• Wired up the same counterapi pattern already used for the page-view counter, under a new feature-colorinspire key — it increments by 1 each time a result is generated, shown below the button.

The keyword matcher, the image-to-palette nearest-neighbor matching, and the no-immediate-repeat logic for random draws were all tested in Node before being written into the file; both language versions were syntax-checked.

Added

• Added a second detection path for diagonal composition: it now also fires when 4+ long lines run the same diagonal direction with at most 1 line running the opposite way. This fixes clearly-diagonal pieces (a reclining figure, a branch sweeping across frame) that previously went undetected.

Fixed

• Fixed a false positive on "pattern break": previously it could fire even when no "repeating pattern" was detected at all; now pattern break can only appear alongside an actual repeating pattern.
• Fixed a variable-naming collision that could let diagonal-detection changes accidentally affect the accuracy of radial composition readings.

This pass re-ran all 26 accumulated sample pieces and manually cross-checked each one, which is how the two bugs above were found.