dsh-background-by-model 0.7.9 → 0.7.10

package/CHANGELOG.md

@@ -1,6 +1,14 @@
 # Changelog
 
-> The full release history lives here; the **Recent optimizations** section of [README.md](./README.md#recent-optimizations) lists only the two most recent releases (v0.7.9 / v0.7.8).
+> The full release history lives here; the **Recent optimizations** section of [README.md](./README.md#recent-optimizations) lists only the two most recent releases (v0.7.10 / v0.7.9).
+
+## v0.7.10
+
+- **The festival art's cache is keyed by the release the plugin pins, so an updated plugin is answered by a download instead of by the previous generation's bytes** — the cache lived at `holiday-cache/<asset>`, and `readHolidayAsset` answers from it whenever the file exists and is a valid image (`const cached = await readFileOrNull(holidayCachePath(asset))` then `if (cached !== null && imageMimeOrNull(cached) !== null) return cached`): no version comparison, no network, no log line. Pointing a build at a new release of the assets repository therefore changed the URL it *would* have fetched and changed nothing about what it returned, because the old bytes were still sitting at `mid-autumn.webp`. That is precisely the SILENT failure the pinned reference exists to prevent — art that stops updating on a day nobody is watching — one layer below the reference, where pinning cannot reach. `holidayCachePath` now carries the tag (`holiday-cache/v0.1.0/mid-autumn.webp`), so the release is part of the file's IDENTITY: a build pinning a new tag looks in a directory nobody has written to, the first holiday after an update is a cache MISS, and `ls holiday-cache/` answers "which generations are on this disk".
+
+- **A superseded generation is dropped, after the new art is on disk and never before it.** A versioned key on its own would leave every generation lying there forever — about 730 KB each (228 KB + 500 KB), including the assets a later release stops shipping at all, which nothing else would ever remove. `pruneHolidayCache` runs at the END of a successful cache write, after the rename: a pinned release that turns out to be unreachable must not cost a user the art they already had, and a download that failed has not earned the right to delete anything. It walks the cache root and removes DIRECTORY entries whose name is not this build's tag — the entries it compares are generation names, not asset names, so the generation in use cannot be touched however the comparison goes — and it is best-effort, because an unprunable cache is a cache with one extra directory in it and failing a festival over that would be the tail wagging the dog. `resetStore` still leaves every generation alone: that cache is not the user's configuration, and pruning is not its job.
+
+- **Gates** — `check:node` had the tag written into two URL assertions and the cache path written as a third literal, so the suite now defines one `REF` that both assertions and `GEN = join(CACHE, REF)` read. Three assertions are new: art missing from the pinned generation is re-downloaded rather than read stale, a planted `v0.0.9` directory is gone once the new art has landed, and the generation in use still holds both assets afterwards. The corrupt-cache-entry, dead-mirror-fallback and raced-cold-read sections were repointed at the generation directory and pass unchanged. `pnpm test` is ten checks, all passing, and `lib/` is that run's output.
 
 ## v0.7.9
 

package/CHANGELOG.zh.md

@@ -1,6 +1,14 @@
 # 更新日志
 
-> 这里保留完整的历史记录；[README.zh.md](./README.zh.md#近期优化) 的「近期优化」一节只列最近两个版本（v0.7.9 / v0.7.8）。
+> 这里保留完整的历史记录；[README.zh.md](./README.zh.md#近期优化) 的「近期优化」一节只列最近两个版本（v0.7.10 / v0.7.9）。
+
+## v0.7.10
+
+- **节日画的缓存现在按插件所钉的发行版做键，于是更新过的插件拿到的是重新下载，而不是上一代的字节** —— 缓存此前住在 `holiday-cache/<asset>`，而 `readHolidayAsset` 只要那个文件存在且是一张合法图片就直接拿它作答（`const cached = await readFileOrNull(holidayCachePath(asset))`，然后 `if (cached !== null && imageMimeOrNull(cached) !== null) return cached`）：不比版本、不联网、不留日志。于是把一个构建指向 assets 仓库的新发行版，改变的是它**本会去取**的 URL，而它**实际返回**的东西一点没变 —— 因为那份旧字节仍然躺在 `mid-autumn.webp` 上。这正是「钉死引用」本身要防的那种**静默**失效（画在没人盯着的那一天停止更新），只是发生在引用之下一层，钉引用够不到的地方。现在 `holidayCachePath` 带上了 tag（`holiday-cache/v0.1.0/mid-autumn.webp`），发行版成了文件**身份**的一部分：钉了新 tag 的构建看的是一个没人写过的目录，更新之后遇到的第一个节日是一次缓存未命中，而 `ls holiday-cache/` 能回答「这张盘上留着哪几代」。
+
+- **被取代的那一代会被清掉，而且是在新图落盘之后、绝不在之前。** 光有带版本的键会让每一代都永远留在那里 —— 每代约 730 KB（228 KB + 500 KB），还包括后续发行版已经彻底不再发布的素材，而那些东西没有别的机制会去删。`pruneHolidayCache` 在一次成功的缓存写入**末尾**、rename 之后运行：一个取不到的发行版不该让用户丢掉他已经有的画，而一次失败的下载也没有资格去删任何东西。它遍历缓存根目录，删掉名字不是本构建 tag 的**目录**条目 —— 它比较的是代号而不是素材名，所以正在用的那一代无论比较结果如何都不可能被碰到 —— 而且它是尽力而为：一个清不掉的缓存只是多了一个目录，为这件事让一个节日失败是本末倒置。`resetStore` 依然不动任何一代：那份缓存不是用户的配置，清理也不是它的职责。
+
+- **门禁** —— `check:node` 此前把 tag 写在两条 URL 断言里、缓存路径又是第三个字面量，现在测试脚本定义了一个 `REF`，两条断言和 `GEN = join(CACHE, REF)` 都读它。新增三条断言：所钉那一代里缺失的素材会被重新下载而不是读到旧的、种下的 `v0.0.9` 目录在新图落盘后消失、以及随后正在用的那一代里两张素材都还在。损坏缓存条目、镜像回落、并发冷读取那几节改成指向代号目录，结果不变。`pnpm test` 十道全过，`lib/` 就是那一次运行的产物。
 
 ## v0.7.9
 

package/README.md

@@ -120,8 +120,8 @@ A small one, not a settings block: **one switch on the Config page and nothing e
 - **National Day** — October 1–7.
 - **Mid-Autumn wins the overlap** — the two do collide (2025-10-06 was both); the single precise day is the better answer.
 - **A holiday with no image falls through** to the model rules rather than blanking the wallpaper.
-- **The art is downloaded, not shipped** — a compressed wallpaper per holiday (228 KB and 500 KB at native resolution, against 8.37 MB of source art), hosted in the project's [assets repository](https://github.com/HarmlessFunny/assets) rather than in the package: they were two thirds of the tarball for a picture that is looked at one day a year. The download is gated on the **calendar**, so a profile that is not on a holiday fetches nothing at all, and the first read on the day caches the file under `holiday-cache/` — every boot after that, and every year after that, is answered from disk and needs no network. The reference is pinned to a **release tag** (`v0.1.0`), never a branch, so reshuffling the repository's screenshots cannot break a shipped festival; a `jsDelivr` mirror is tried first and `raw.githubusercontent.com` second, and if both are unreachable the holiday simply does not take over — the model rules keep painting.
-- **Neither the art nor the palette is swappable, by design** — an easter egg that asks to be configured is not an easter egg. A holiday slot is read-only: a file dropped into `modelbg-h-midautumn` / `modelbg-h-nationalday` is ignored by every read, and the plugin refuses to write, fetch into or delete those slots. The download lands in `holiday-cache/`, keyed by asset name and never by slot, so the store still never holds a `modelbg-h-*` file. The theme colors are constants in the holiday definition and are forced on every read — on the holiday **and on its image**, since a theme color lives on the image now — so a hand-edited `color` in `theme-config.json` cannot repaint a festival either. Its layout mode is the definition's **Fill** for the same reason (the art is full-bleed, and a stale `fit` would letterbox it), while the **framing** next to it stays yours. Theme exports therefore carry your rules only.
+- **The art is downloaded, not shipped** — a compressed wallpaper per holiday (228 KB and 500 KB at native resolution, against 8.37 MB of source art), hosted in the project's [assets repository](https://github.com/HarmlessFunny/assets) rather than in the package: they were two thirds of the tarball for a picture that is looked at one day a year. The download is gated on the **calendar**, so a profile that is not on a holiday fetches nothing at all, and the first read on the day caches the file under `holiday-cache/<tag>/` — **keyed by the release the plugin pins**, so a build that pins a new tag is answered by a fresh download instead of by the previous generation's bytes, and every boot after that, and every year after that, is answered from disk and needs no network. The reference is pinned to a **release tag** (`v0.1.0`), never a branch, so reshuffling the repository's screenshots cannot break a shipped festival; a `jsDelivr` mirror is tried first and `raw.githubusercontent.com` second, and if both are unreachable the holiday simply does not take over — the model rules keep painting.
+- **Neither the art nor the palette is swappable, by design** — an easter egg that asks to be configured is not an easter egg. A holiday slot is read-only: a file dropped into `modelbg-h-midautumn` / `modelbg-h-nationalday` is ignored by every read, and the plugin refuses to write, fetch into or delete those slots. The download lands in `holiday-cache/<tag>/`, named after the asset and never after a slot, so the store still never holds a `modelbg-h-*` file. The theme colors are constants in the holiday definition and are forced on every read — on the holiday **and on its image**, since a theme color lives on the image now — so a hand-edited `color` in `theme-config.json` cannot repaint a festival either. Its layout mode is the definition's **Fill** for the same reason (the art is full-bleed, and a stale `fit` would letterbox it), while the **framing** next to it stays yours. Theme exports therefore carry your rules only.
 - **Turning it on** — the switch on the Config page, or `"holidays": { "enabled": true }` in the same file. Only a real `true` counts: the feature is on because a configuration says so, not because nothing has been said. The per-holiday switches inside it (中秋 / 国庆) are a different thing and still default to on — they are details of a feature that is already switched on.
 
 ### Per-rule and per-image settings
@@ -188,7 +188,7 @@ Data directory: `~/.dsh/.dsh-background-by-model-data/` (Windows: `C:\Users\<you
 | --- | --- |
 | `theme-config.json` | The rule list (each rule with its image list and rotation) plus the global Interface settings, the global switch effect and the holiday overrides |
 | `modelbg-<slot>` | One image, stored as raw bytes without a file extension — one file per image, so a rule with three images owns three slots |
-| `holiday-cache/<asset>` | Festival art downloaded from the CDN, one file per holiday (`mid-autumn.webp`, `national-day.webp`). Keyed by **asset name**, deliberately outside the slot namespace — a holiday slot still never holds a file, which is what keeps it read-only. Delete a file here and it is simply re-downloaded the next time that holiday is in force |
+| `holiday-cache/<tag>/<asset>` | Festival art downloaded from the CDN, one file per holiday (`mid-autumn.webp`, `national-day.webp`) inside a directory named after the assets release this build pins (`v0.1.0`). That directory is the point: a cache keyed by asset name alone would answer an updated plugin with the **previous** generation's bytes and say nothing about it, so the release is part of the path and a build pinning a new tag starts from an empty directory. Inside it the file is named after the **asset**, deliberately outside the slot namespace — a holiday slot still never holds a file, which is what keeps it read-only. Superseded generations are pruned once the new art is on disk. Delete a file here and it is simply re-downloaded the next time that holiday is in force |
 
 ## Installation
 
@@ -251,7 +251,7 @@ After changing anything under `src/`, run `pnpm run bundle` again — the mounte
 - **`check:rules`** — the two decisions behind the tab strip, which are pure and live in `src/client/rules-view.ts` (which rule the editor shows, and where a deletion lands — its neighbour, never rule 1), plus the page structure the layout rests on: one editor drawn, no stacked per-card container, no per-card open state, a strip that can neither wrap nor grow, and its scrollbar block staying scoped to the plugin's own classes.
 - **`check:preset`** — the two decisions behind the recommended profile's transfers (`src/preset.ts`: which failures are worth another round, and what a failed attempt leaves for the next one), plus the two halves actually using them. Each of the six failure reasons is judged on its own (the three terminal ones are the three that can be decided locally and cheaply, and a timeout is not retried because both mirrors were already asked inside the same round); the wait grows and stays bounded; two key orders of one value are one revision while the image list's ORDER counts (it is the rotation); and a resume hands back a copy rather than the held record itself. The wiring is pinned with it: the node half looping over the round budget and asking the policy, the client sizing its pool by what is missing rather than by the whole slot list, holding what arrived on failure and clearing it on success, and applying the size cap **before** a file is kept.
 - **`check:ui`** — every `t('…')` key exists in both dictionaries, both dictionaries carry the same key set, every `dab-…` class has a rule in the stylesheet, no dictionary entry has gone unreferenced, and every `var(--dsw-…)` the UI reads is a token the installed theme actually publishes — an undefined custom property drops the declaration and the property quietly falls back to an inherited value, which is how a placeholder once painted in the ink of real text (with the theme bundle absent, that section skips with a notice).
-- **`check:node`** — the built node half driven through its real RPC handler with `DSH_HOME` redirected to a throwaway directory, and `globalThis.fetch` replaced by a stub — the festival art lives on a CDN now, and this suite must not depend on one being up, nor assert today's bytes of a file it does not own. It covers: the holiday block through the shared sanitizer; nothing fetched merely for loading the plugin or reading the config; the art fetched on demand and cached under `holiday-cache/` keyed by **asset** rather than slot; a second read answered from disk with no second transfer; two simultaneous cold reads sharing one transfer; a dead first mirror falling through to the second with both mirrors read at the same release tag; every mirror down yielding "no bytes" rather than a throw, with nothing cached from a failed fetch; a mirror answering 200 with a non-image rejected rather than cached; a corrupt cache entry discarded and re-downloaded; a holiday slot refusing writes, deletes and URL fetches while a file dropped into it is ignored and no `modelbg-h-*` file is ever written; an ordinary rule slot still writing, listing and deleting; a hand-edited config unable to redirect a holiday at another rule's image; the per-image color shape — the read-time lift in both directions (an absent key inherits the rule's color, an explicit `null` does not), independent colors through the file, and a `null` written as a key rather than omitted; and the global switch effect — the shipped default for a config that predates it, a chosen effect/easing/unified duration surviving the round trip, hostile values falling back while an absurd duration clamps, a `0` duration kept rather than defaulted, and drift inside `transition` reported instead of silently dropped. The recommended profile is covered end to end from the node half's side: the manifest and the config fetched in that order and sanitized on arrival, the request set DERIVED from the rules with the holiday slots dropped from it, one request per wallpaper with the bytes sniffed rather than trusted from a header, a dead jsDelivr falling through to raw for a single file, the same file **asked again** when both mirrors stumble at once and arriving on that third request (three entries in the log), a file the host does not have still costing exactly two requests (one per mirror, no waiting), a permanently broken mirror costing exactly six (three rounds over two mirrors — the retry has a budget), and every way it can fail reported as a value rather than a throw — every mirror down, a mirror answering 500, a 200 carrying HTML, a manifest whose version is not a number, a profile written for a newer shape refused before a single wallpaper is asked for, a config that is not JSON, a config that is a JSON array, a config naming more images than the cap, and a mirror announcing an oversized body; plus the strict half (a named wallpaper the host does not have is an error, not a short profile), the two naming refusals that happen before any request at all (a holiday slot, and a slot name that is not a slot), and the reset: the config and every rule image gone, `holiday-cache/` untouched, a read afterwards yielding the same one blank rule rather than an error, and the profile then writing into the emptied store — with the fresh-install rule itself pinned first: a data directory with no config file reads as exactly one blank rule (`r1`, no match string, no colour, no image), while a rule list written explicitly empty still reads back empty, because those are two different answers.
+- **`check:node`** — the built node half driven through its real RPC handler with `DSH_HOME` redirected to a throwaway directory, and `globalThis.fetch` replaced by a stub — the festival art lives on a CDN now, and this suite must not depend on one being up, nor assert today's bytes of a file it does not own. It covers: the holiday block through the shared sanitizer; nothing fetched merely for loading the plugin or reading the config; the art fetched on demand and cached under `holiday-cache/<pinned tag>/` — keyed by the **release** and by **asset**, never by slot, with a superseded generation dropped once the new art is on disk; a second read answered from disk with no second transfer; two simultaneous cold reads sharing one transfer; a dead first mirror falling through to the second with both mirrors read at the same release tag; every mirror down yielding "no bytes" rather than a throw, with nothing cached from a failed fetch; a mirror answering 200 with a non-image rejected rather than cached; a corrupt cache entry discarded and re-downloaded; a holiday slot refusing writes, deletes and URL fetches while a file dropped into it is ignored and no `modelbg-h-*` file is ever written; an ordinary rule slot still writing, listing and deleting; a hand-edited config unable to redirect a holiday at another rule's image; the per-image color shape — the read-time lift in both directions (an absent key inherits the rule's color, an explicit `null` does not), independent colors through the file, and a `null` written as a key rather than omitted; and the global switch effect — the shipped default for a config that predates it, a chosen effect/easing/unified duration surviving the round trip, hostile values falling back while an absurd duration clamps, a `0` duration kept rather than defaulted, and drift inside `transition` reported instead of silently dropped. The recommended profile is covered end to end from the node half's side: the manifest and the config fetched in that order and sanitized on arrival, the request set DERIVED from the rules with the holiday slots dropped from it, one request per wallpaper with the bytes sniffed rather than trusted from a header, a dead jsDelivr falling through to raw for a single file, the same file **asked again** when both mirrors stumble at once and arriving on that third request (three entries in the log), a file the host does not have still costing exactly two requests (one per mirror, no waiting), a permanently broken mirror costing exactly six (three rounds over two mirrors — the retry has a budget), and every way it can fail reported as a value rather than a throw — every mirror down, a mirror answering 500, a 200 carrying HTML, a manifest whose version is not a number, a profile written for a newer shape refused before a single wallpaper is asked for, a config that is not JSON, a config that is a JSON array, a config naming more images than the cap, and a mirror announcing an oversized body; plus the strict half (a named wallpaper the host does not have is an error, not a short profile), the two naming refusals that happen before any request at all (a holiday slot, and a slot name that is not a slot), and the reset: the config and every rule image gone, `holiday-cache/` untouched, a read afterwards yielding the same one blank rule rather than an error, and the profile then writing into the emptied store — with the fresh-install rule itself pinned first: a data directory with no config file reads as exactly one blank rule (`r1`, no match string, no colour, no image), while a rule list written explicitly empty still reads back empty, because those are two different answers.
 
 ## FAQ
 
@@ -302,7 +302,13 @@ Two things can hold it, and both are on the card: the effect is **Instant**, or
 
 ## Recent Optimizations
 
-> Only the two most recent releases are listed here; older ones (v0.7.7 and earlier) live in [CHANGELOG.md](./CHANGELOG.md).
+> Only the two most recent releases are listed here; older ones (v0.7.8 and earlier) live in [CHANGELOG.md](./CHANGELOG.md).
+
+### v0.7.10
+
+- **The festival art's cache is keyed by the pinned release, so an updated plugin re-downloads instead of being answered with the old bytes** — the cache lived at `holiday-cache/<asset>`, which made the file's identity independent of the bytes behind it: point the plugin at a new release of the assets repository and `readHolidayAsset` still found `mid-autumn.webp` on disk, returned it, and said nothing. That is the same silent failure the pinned reference exists to prevent, one layer down — and it is invisible by construction, on a day nobody is watching. The cache is now `holiday-cache/<tag>/<asset>` (`holiday-cache/v0.1.0/mid-autumn.webp`), so the tag is part of the path and a build pinning a new tag starts from a directory nobody has written to. `ls holiday-cache/` now answers "which generations are on this disk".
+- **The generation it supersedes is dropped, after the new art is on disk and never before it.** A versioned key would otherwise leave every previous generation lying there forever (about 730 KB each), including the assets a later release stops shipping at all. Pruning happens at the end of a successful cache write — the bytes are already in hand and renamed into place — so a pinned release that turns out to be unreachable cannot cost a user the art they already had. It removes **directory** entries and only those whose name is not this build's tag, so the generation in use cannot be touched however the comparison goes.
+- **`check:node` pins both halves of it.** The suite's `REF` is now one value used by the two URL assertions *and* by the cache path, and three assertions were added: art missing from the pinned generation is re-downloaded rather than read stale, a planted `v0.0.9` directory is gone once the new art lands, and the generation in use is left exactly where it was.
 
 ### v0.7.9
 
@@ -311,12 +317,6 @@ Two things can hold it, and both are on the card: the effect is **Instant**, or
 - **The recommended profile deliberately gives up "fix it without a release"** — it used to follow a branch so a bad recommendation could be corrected in the assets repository instead of waiting for a plugin release, and the reasoning was that a profile fails *loudly* (you are looking at the button) where a holiday fails silently, so immutability bought nothing there. What it cost was a plugin whose bytes could change under you between the install and the click; with one tag covering the screenshots, the art and the profile, that is no longer true of any of them. Correcting one JSON file now costs a tag in `assets`, a value in `src/index.ts` and a release on npm. Note what a tag is **not**: it can be re-pointed, and jsDelivr serves a version reference as immutable, so re-pointing one means a purge request rather than a push — which is why `presetKey` survives this change, since an edge that has not been purged can still answer with the previous bytes of the same URL, and a press that resumes must not glue two revisions together.
 - **`check:node` pins the tag on both features.** Its `revision()` helper used to look for a 40-character SHA, which cannot see a tag at all; it now lifts the reference out beside the repository name (`HarmlessFunny/assets@<ref>` in the jsDelivr shape, `HarmlessFunny/assets/<ref>/` in the raw one) so the two mirrors can still be compared, and the two assertions that used to read "a commit, never a branch" and "follows the branch" now both require `v0.1.0`.
 
-### v0.7.8
-
-- **A fresh install opens on one blank rule instead of the "no rules yet" screen** — with no config file at all the read used to produce an empty rule list, so the first thing a new user saw was "No rules yet — use 'Add rule' in the tab strip above": the panel telling them to press a control it had not given them. "The `rules` key is not there" (nothing has ever been configured) and "the key is there and the array is empty" (the user deleted every rule) are two different answers now, and only the first is filled in with one blank rule (id `r1`, so it is rule 1, the fallback, and the tab the panel selects); the second stays empty, so deleting the last rule is not undone by the next read. The rule is deliberately blank rather than half-built: with no picture and no theme colour of its own, matching skips it until something is filled in, instead of letting an empty background take the fallback away from rule 1.
-- **The placeholder no longer looks like real text** — its `::placeholder` colour named `--dsw-alias-label-quaternary`, which is defined NOWHERE in this harness's theme: an undefined custom property drops the declaration at computed-value time, and `color` is inherited, so the placeholder fell back to the input's own `label-primary` — the brightest ink in the panel, inside a box nobody had typed in. All three inputs that carry one (the rule's match string, the image URL, the rotation's custom seconds) now read `--dsw-alias-label-caption`, the weak token the plugin already uses for input placeholders; two more dead names went with it (`.dab-pick-n`'s `--dsw-mono` was missing a letter — the plugin's own variable is `--dab-mono` — and `.dab-brand-tag`'s carried a fallback, so dropping it changes no colour).
-- **The recommended profile retries per file, and a failed attempt keeps what it fetched** — a file used to get exactly two chances (one per mirror), so with the first few wallpapers down and the next hitting a 502 or a CDN hiccup the whole profile failed and pressing the button again started from zero. Every file now gets up to three rounds, 400 ms and then 800 ms apart; three failures are not retried — "the file is not there" (the same answer twice), "the bytes are over the cap" (the same bytes again), and "a full minute of silence" (both mirrors were already asked inside that same round, and this is what keeps the worst case at three minutes rather than six) — while everything else gets another round. What did arrive stays in hand, so the next press fetches only what is missing, as long as the config it re-fetches is **the same value** as the one those bytes came from (the profile is served from a branch, and two different pushes must not be glued together). The confirmation, the button's copy and where the progress sits are untouched; **so is the strictness** — the first file that cannot be had still fails the whole attempt and still names the slot.
-
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/chart?repos=HarmlessFunny/dsh-background-by-model&type=timeline&legend=bottom-right)](https://www.star-history.com/?repos=HarmlessFunny%2Fdsh-background-by-model&type=timeline&legend=bottom-right)

package/README.zh.md

@@ -120,8 +120,8 @@
 - **国庆节** — 公历 10 月 1–7 日。
 - **两者重叠时中秋优先** — 它们确实会撞（2025-10-06 就是同一天），一个精确的单日比七天窗口里的某一天更值得采信。
 - **没有图的节日会回落到模型规则**，而不是把背景刷空。
-- **画是下载来的，不是随包发的** — 每个节日一张压缩过的壁纸（原生分辨率下 228 KB 与 500 KB，原图共 8.37 MB），放在项目自己的 [assets 仓库](https://github.com/HarmlessFunny/assets)里，而不是打进包里 —— 它们占了整个 tarball 的三分之二，而这张画一年只被看一天。下载由**日历**把关：不在节日的配置一点流量都不会有；当天第一次读取会把它缓存到 `holiday-cache/`，此后每一次启动、往后每一年，都直接读磁盘、不再联网。引用钉死在一个**发行 tag**（`v0.1.0`）上而不是分支，所以你以后整理仓库里的截图不可能弄坏一个已经发出去的节日；先走 `jsDelivr`、失败再走 `raw.githubusercontent.com`，两个都不通时这个节日就是不接管 —— 模型规则继续画，不会白屏。
-- **图片和配色都不可替换，这是刻意的** — 一个要求用户去配置的彩蛋就不是彩蛋了。节日槽位是只读的：往 `modelbg-h-midautumn` / `modelbg-h-nationalday` 里放文件不会被任何一次读取采用，插件也拒绝写入、抓取和删除这两个槽位。下载落进 `holiday-cache/`，按**素材名**而不是槽位命名，所以数据目录里依然永远不会出现 `modelbg-h-*` 文件。主题色是节日定义里的常量、每次读取都强制覆盖——**节日条目和它的图片两处都覆盖**，所以手改 `theme-config.json` 里的 `color` 也换不掉节日配色。布局模式同理固定为定义里的**填充**（这些画本来就是满幅的，一个残留的 `fit` 会给它加黑边），而旁边的**取景**仍然是你自己的。导出的主题文件只包含你自己的规则。
+- **画是下载来的，不是随包发的** — 每个节日一张压缩过的壁纸（原生分辨率下 228 KB 与 500 KB，原图共 8.37 MB），放在项目自己的 [assets 仓库](https://github.com/HarmlessFunny/assets)里，而不是打进包里 —— 它们占了整个 tarball 的三分之二，而这张画一年只被看一天。下载由**日历**把关：不在节日的配置一点流量都不会有；当天第一次读取会把它缓存到 `holiday-cache/<tag>/` —— 路径里带**插件所钉的那个发行版**，于是一个换成新 tag 的构建拿到的是重新下载，而不是上一代的字节 —— 此后每一次启动、往后每一年，都直接读磁盘、不再联网。引用钉死在一个**发行 tag**（`v0.1.0`）上而不是分支，所以你以后整理仓库里的截图不可能弄坏一个已经发出去的节日；先走 `jsDelivr`、失败再走 `raw.githubusercontent.com`，两个都不通时这个节日就是不接管 —— 模型规则继续画，不会白屏。
+- **图片和配色都不可替换，这是刻意的** — 一个要求用户去配置的彩蛋就不是彩蛋了。节日槽位是只读的：往 `modelbg-h-midautumn` / `modelbg-h-nationalday` 里放文件不会被任何一次读取采用，插件也拒绝写入、抓取和删除这两个槽位。下载落进 `holiday-cache/<tag>/`，文件名按**素材名**而不是槽位，所以数据目录里依然永远不会出现 `modelbg-h-*` 文件。主题色是节日定义里的常量、每次读取都强制覆盖——**节日条目和它的图片两处都覆盖**，所以手改 `theme-config.json` 里的 `color` 也换不掉节日配色。布局模式同理固定为定义里的**填充**（这些画本来就是满幅的，一个残留的 `fit` 会给它加黑边），而旁边的**取景**仍然是你自己的。导出的主题文件只包含你自己的规则。
 - **想要它** — 打开「配置」页那个开关，或者在同一个文件里写 `"holidays": { "enabled": true }`。只有真正的 `true` 算数：这个功能是「配置说了要」才开的，不是「没人说话」就默认开。它内部每个节日（中秋 / 国庆）各自那个开关是另一回事，仍然默认开着 —— 那是一个已经打开的功能内部的细节。
 
 ### 规则与图片各自的设置
@@ -188,7 +188,7 @@ preset/
 | --- | --- |
 | `theme-config.json` | 规则列表（每条规则含图片列表与轮换设置）+ 全局「界面」设置 + 全局切换效果 + 节日覆盖配置 |
 | `modelbg-<slot>` | 一张图片，保存为原始字节，无扩展名 —— 一张图一个文件，所以一条有三张图的规则占三个槽位 |
-| `holiday-cache/<asset>` | 从 CDN 下载的节日画，一个节日一个文件（`mid-autumn.webp`、`national-day.webp`）。按**素材名**命名，刻意放在槽位命名空间之外 —— 节日槽位里依然永远不会有文件，这正是它只读的原因。删掉这里的文件也没关系，下次那个节日生效时会重新下载 |
+| `holiday-cache/<tag>/<asset>` | 从 CDN 下载的节日画，一个节日一个文件（`mid-autumn.webp`、`national-day.webp`），装在一个以**插件所钉的 assets 发行版**命名的目录里（`v0.1.0`）。这个目录是关键：只按素材名做键的缓存会把上一代的字节交给一个已经更新过的插件、而且什么都不说，所以发行版被放进路径里，换成新 tag 的构建从一个没人写过的目录开始。目录里的文件名仍然是**素材名**，刻意放在槽位命名空间之外 —— 节日槽位里依然永远不会有文件，这正是它只读的原因。被取代的旧代会在新图落盘之后被清掉。删掉这里的文件也没关系，下次那个节日生效时会重新下载 |
 
 ## 安装
 
@@ -251,7 +251,7 @@ pnpm run typecheck
 - **`check:rules`** — 选项卡背后那两个纯判定（在 `src/client/rules-view.ts`：编辑器显示哪条规则、删除之后落在哪——落在邻居上，永远不是规则 1），以及布局所依赖的页面结构：只画一个编辑器、没有旧的堆叠容器、没有逐卡片的展开状态、胶片条既不换行也不长高，以及它的滚动条样式块仍然被限定在插件自己的类上。
 - **`check:preset`** — 推荐配置那几次传输背后的两个判定（`src/preset.ts`：哪些失败值得再来一轮、失败的那一次给下一次留下什么），以及两侧**真的在用**它们：六个失败原因逐个判定（三个终态正是能在本地廉价判定的那三个；超时不重试是因为同一轮里两个镜像都已经问过）、等待指数增长且有上限、同一个值的不同键顺序给出同一个「同一份配置」而图片列表的顺序**算数**（它就是轮换顺序）、续传拿到的必须是副本而不是那条记录本身 —— 再加上接线（node 半边循环轮次并按策略决定、客户端按「还缺哪些」而不是整个槽位表铺并发池、失败留字节、成功清掉、体积上限在收下文件之前判）。
 - **`check:ui`** — 每个 `t('…')` 键在中英两本字典里都存在、两本字典键集完全一致、每个 `dab-…` 类在样式表里都有规则、没有字典条目变成没人引用的死键，以及每一个 `var(--dsw-…)` 都必须是已安装主题确实发布过的 token —— 未定义的自定义属性会让声明被丢掉、属性悄悄回落到继承值，placeholder 因此曾经画成正文字的墨色（读不到主题时这一节跳过并说明）。
-- **`check:node`** — 把构建好的 node 半边通过它真实的 RPC 处理器跑一遍，`DSH_HOME` 指向临时目录，并把 `globalThis.fetch` 换成桩 —— 节日画现在住在 CDN 上，这套校验不能依赖某个 CDN 是活的，也不该去断言一个不属于自己的文件今天有多少字节。它覆盖：节日配置块穿过共享的 sanitizer；仅仅加载插件或读取配置不会发起任何下载；画是按需下载并缓存到 `holiday-cache/`、按**素材名**而不是槽位命名；第二次读取直接读盘、不再传输；两次并发的冷读取共用一次传输；第一个镜像不通时回落到第二个，且两个镜像读的是同一个发行 tag；所有镜像都不通时答案是「没有字节」而不是抛异常，失败的抓取什么都不缓存；镜像返回 200 但内容不是图片时被拒绝而不是被缓存；损坏的缓存条目被丢弃并重新下载；节日槽位拒绝写入 / 删除 / 抓取而丢进去的文件被忽略、且数据目录里永远不会出现 `modelbg-h-*` 文件；普通规则槽位照常写入和删除；手改过的配置无法把节日指向别的规则的图片。它还覆盖逐图主题色这套形状：读取时抬升的两个方向（缺键继承规则色、显式 `null` 不继承）、颜色经过文件往返后各自独立，以及 `null` 是写出键而不是省略。切换效果同样在覆盖范围内：旧配置拿到的出厂默认、选定的效果 / 缓动 / 统一时长能原样穿过文件往返、非法值回退而夸张时长被夹紧、时长为 `0` 被保留而不是回退成默认，以及 `transition` 内部的字段漂移会被上报而不是悄悄丢掉。推荐配置从 node 半边这一侧被整条覆盖：manifest 与 config 按序下载并当场穿过 sanitizer、请求清单由规则推导且把节日槽位剔掉、一张壁纸一个请求且类型是嗅出来的而不是信 header、jsDelivr 对单个文件死了会回落到 raw、两个镜像同时抖一下时同一个文件会被**再问一遍**并最终拿到（三个请求，日志里看得见）、托管处没有的文件仍然只花两次请求（每镜像一次、不等待）、永久 500 的镜像总共只花六次请求（三轮 × 两个镜像 —— 重试是有预算的），以及它每一种失败都是返回值而不是抛异常 —— 所有镜像都不通、镜像回 500、回 200 但内容是 HTML、manifest 的版本不是数字、写给更新形状的配置在请求第一张壁纸之前就被拒绝、config 不是 JSON、config 是 JSON 数组、config 列出的图片数超限、镜像谎报一个超大的长度；再加上严格的那一半（config 点名而托管处没有的壁纸是一个错误，而不是一份短了的配置）、两次在任何请求之前就发生的拒绝（节日槽位、以及一个根本不是槽位的名字），以及新装那条规则：空数据目录读出来正好一条空白规则（`r1`、无匹配串、无颜色、无图），而显式写空的规则列表读回仍然为空 —— 这两者是不同的答案；清空那一步：配置和每一张规则图片都没了、`holiday-cache/` 原封不动、清空后再读拿到的是同一条空白规则而不是报错，以及随后能往空目录里写入这份配置。
+- **`check:node`** — 把构建好的 node 半边通过它真实的 RPC 处理器跑一遍，`DSH_HOME` 指向临时目录，并把 `globalThis.fetch` 换成桩 —— 节日画现在住在 CDN 上，这套校验不能依赖某个 CDN 是活的，也不该去断言一个不属于自己的文件今天有多少字节。它覆盖：节日配置块穿过共享的 sanitizer；仅仅加载插件或读取配置不会发起任何下载；画是按需下载并缓存到 `holiday-cache/<所钉 tag>/` —— 按**发行版**和**素材名**而不是槽位命名，被取代的旧代会在新图落盘之后被清掉；第二次读取直接读盘、不再传输；两次并发的冷读取共用一次传输；第一个镜像不通时回落到第二个，且两个镜像读的是同一个发行 tag；所有镜像都不通时答案是「没有字节」而不是抛异常，失败的抓取什么都不缓存；镜像返回 200 但内容不是图片时被拒绝而不是被缓存；损坏的缓存条目被丢弃并重新下载；节日槽位拒绝写入 / 删除 / 抓取而丢进去的文件被忽略、且数据目录里永远不会出现 `modelbg-h-*` 文件；普通规则槽位照常写入和删除；手改过的配置无法把节日指向别的规则的图片。它还覆盖逐图主题色这套形状：读取时抬升的两个方向（缺键继承规则色、显式 `null` 不继承）、颜色经过文件往返后各自独立，以及 `null` 是写出键而不是省略。切换效果同样在覆盖范围内：旧配置拿到的出厂默认、选定的效果 / 缓动 / 统一时长能原样穿过文件往返、非法值回退而夸张时长被夹紧、时长为 `0` 被保留而不是回退成默认，以及 `transition` 内部的字段漂移会被上报而不是悄悄丢掉。推荐配置从 node 半边这一侧被整条覆盖：manifest 与 config 按序下载并当场穿过 sanitizer、请求清单由规则推导且把节日槽位剔掉、一张壁纸一个请求且类型是嗅出来的而不是信 header、jsDelivr 对单个文件死了会回落到 raw、两个镜像同时抖一下时同一个文件会被**再问一遍**并最终拿到（三个请求，日志里看得见）、托管处没有的文件仍然只花两次请求（每镜像一次、不等待）、永久 500 的镜像总共只花六次请求（三轮 × 两个镜像 —— 重试是有预算的），以及它每一种失败都是返回值而不是抛异常 —— 所有镜像都不通、镜像回 500、回 200 但内容是 HTML、manifest 的版本不是数字、写给更新形状的配置在请求第一张壁纸之前就被拒绝、config 不是 JSON、config 是 JSON 数组、config 列出的图片数超限、镜像谎报一个超大的长度；再加上严格的那一半（config 点名而托管处没有的壁纸是一个错误，而不是一份短了的配置）、两次在任何请求之前就发生的拒绝（节日槽位、以及一个根本不是槽位的名字），以及新装那条规则：空数据目录读出来正好一条空白规则（`r1`、无匹配串、无颜色、无图），而显式写空的规则列表读回仍然为空 —— 这两者是不同的答案；清空那一步：配置和每一张规则图片都没了、`holiday-cache/` 原封不动、清空后再读拿到的是同一条空白规则而不是报错，以及随后能往空目录里写入这份配置。
 
 ## 常见问题
 
@@ -302,7 +302,13 @@ pnpm run typecheck
 
 ## 近期优化
 
-> 这里只列最近两个版本；更早的版本（v0.7.7 及以前）见 [CHANGELOG.zh.md](./CHANGELOG.zh.md)。
+> 这里只列最近两个版本；更早的版本（v0.7.8 及以前）见 [CHANGELOG.zh.md](./CHANGELOG.zh.md)。
+
+### v0.7.10
+
+- **节日画的缓存现在按所钉的发行版做键，于是更新过的插件会重新下载，而不是被告知旧的字节就是答案** —— 缓存此前住在 `holiday-cache/<asset>`，这让「文件的身份」和「它背后的字节」脱了钩：把插件指向 assets 仓库的一个新发行版，`readHolidayAsset` 依然能在磁盘上找到 `mid-autumn.webp`，把它返回，然后什么都不说。这正是「钉死引用」本身要防的那种静默失效，只是低了一层 —— 而且它是构造上不可见的，发生在没人盯着的那一天。现在缓存是 `holiday-cache/<tag>/<asset>`（`holiday-cache/v0.1.0/mid-autumn.webp`），tag 进了路径，于是一个钉了新 tag 的构建从一个没人写过的目录开始。`ls holiday-cache/` 现在能回答「这张盘上留着哪几代」。
+- **被取代的那一代会被清掉，而且是在新图落盘之后、绝不在之前。** 带版本的键本来会让每一代都永远留在那里（每代约 730 KB），包括后续发行版已经不再发布的素材。清理发生在一次成功的缓存写入的末尾 —— 那时字节已经在手里、也已经 rename 到位 —— 所以一个取不到的发行版不会让用户丢掉他已经有的画。它删的是**目录**条目、而且只删名字不是本构建 tag 的那些，所以正在用的那一代无论比较结果如何都不可能被碰到。
+- **`check:node` 把这两半都钉住了。** 测试脚本里的 `REF` 现在是一个值，两条 URL 断言**和**缓存路径共用它；另外新增三条断言：所钉那一代里缺失的素材会被重新下载而不是读到旧的、种下的 `v0.0.9` 目录在新图落盘后消失、以及正在用的那一代原封不动。
 
 ### v0.7.9
 
@@ -311,12 +317,6 @@ pnpm run typecheck
 - **推荐配置是刻意放弃「不发版就能修」的** —— 它此前跟着分支，是为了让一个不好的推荐能在 assets 仓库里直接修正、不必等插件发版，理由是推荐配置的失败是**响亮的**（你正盯着那个按钮），而节日的失败是静默的，所以不可变在这里什么都买不到。它买来的代价是：插件的字节可能在装上之后、你点下去之前发生变化；现在一个 tag 同时覆盖截图、节日素材和这份配置，三者都不再有这个问题。改一个 JSON 现在要花一个 `assets` 里的 tag、`src/index.ts` 里的一个值和一次 npm 发版。也要注意 tag **不是**什么：它可以被重新指向，而 jsDelivr 把版本引用按不可变处理，所以重新指向要发一次 purge 请求而不是 push —— 这也是 `presetKey` 在这次改动后仍然保留的原因：一个没清过缓存的边缘节点仍会回同一个 URL 的上一次字节，恢复下载的那一次不能把两个修订拼在一起。
 - **`check:node` 在两个功能上都钉住了这个 tag。** 它的 `revision()` 辅助函数此前找的是 40 位 SHA，根本看不见 tag；现在它从仓库名旁边取出引用（jsDelivr 形状是 `HarmlessFunny/assets@<ref>`，raw 形状是 `HarmlessFunny/assets/<ref>/`），于是两个镜像仍然可以互相比较；两条断言里原来的「a commit, never a branch」和「follows the branch」现在都要求 `v0.1.0`。
 
-### v0.7.8
-
-- **新装打开的就是一条空白规则，而不是「还没有规则」那一屏** —— 此前完全没有配置文件时读出来的是一份空的规则列表，于是第一次装上看到的是那句「还没有规则，点上面的『新增规则』新建一条」：界面在教用户去点一个它还没给出的控件。现在「`rules` 这个键根本不在」（从未配置过）和「键在、数组是空的」（用户把规则删光了）是两个不同的答案，只有前者会种出一条空白规则（id 是 `r1`，所以它就是规则 1、就是兜底、就是面板默认选中的那张卡），后者保持为空 —— 删掉最后一条规则不会被下一次读盘复活。这条规则刻意是「空」而不是「半成品」：它没有图、也没有自己的主题色，所以在被填上东西之前匹配会跳过它，而不是让一个空背景顶掉规则 1 的兜底。
-- **placeholder 不再看起来像真的文字** —— 那条 `::placeholder` 颜色写的 `--dsw-alias-label-quaternary` 在这个 harness 的主题里**一个定义都没有**：未定义的自定义属性会让声明在计算值阶段被丢掉，而 `color` 是继承属性，于是 placeholder 拿到输入框自己的 `label-primary` —— 面板里最亮的墨色，写在一个用户一个字都没打过的框里。三个带 placeholder 的输入（规则匹配串、图片网址、轮换的自定义秒数）现在统一读 `--dsw-alias-label-caption`，也就是插件早就给输入框用的那个弱一档的 token；同一个死名字还有两处跟着修掉（`.dab-pick-n` 的 `--dsw-mono` 少了个字母，插件自己的变量是 `--dab-mono`；`.dab-brand-tag` 的那个本来就带 fallback，去掉它不改变颜色）。
-- **推荐配置的下载会按文件重试，失败的那一次把已下到的部分留给下一次** —— 此前一个文件只有两次机会（两个镜像各一次），前面几张都下好了、下一张碰上一次 502 或者一次 CDN 抽风，整份配置就失败，而再按一次按钮是从零开始。现在每个文件最多三轮、轮间 400 ms 再 800 ms；三种失败不重试 ——「这个文件不在那儿」（再问也是同一个答案）、「字节本身就超限」（下一轮还是这些字节）、「六十秒都没吭声」（两个镜像在同一轮里都已经问过了，而这一条正是把最坏情况钉在三分钟而不是六分钟的东西），其余失败会再问一轮。失败的那一次把已经拿到的字节留在手里，下一次只补缺的那几张 —— 前提是重新取到的配置与那些字节来自的那一份**逐值相同**（这份配置服务在分支上，两次不同的推送不许拼在一起）。确认框、按钮文案、进度读数的位置一处没动；**严格性也没变**：第一个拿不到的文件仍然让整次失败、仍然报出是哪一个。
-
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/chart?repos=HarmlessFunny/dsh-background-by-model&type=timeline&legend=bottom-right)](https://www.star-history.com/?repos=HarmlessFunny%2Fdsh-background-by-model&type=timeline&legend=bottom-right)

package/lib/index.js

@@ -569,8 +569,9 @@ function presetRetryDelayMs(failedAttempt) {
 *   theme-config.json   the ordered rule list + the global interface settings
 *   modelbg-<slot>      one background image per rule (raw bytes, no extension;
 *                       the MIME is sniffed from the magic bytes when served)
-*   holiday-cache/      festival art downloaded from the CDN, one file per
-*                       holiday asset — deliberately NOT a slot (see below)
+*   holiday-cache/<tag>/  festival art downloaded from the CDN — one directory
+*                       per pinned release of the asset repository, one file per
+*                       holiday asset inside it, deliberately NOT a slot (below)
 *
 * Images travel as data URLs over the RPC channel — there is no separate HTTP
 * image route. A legacy single-wallpaper / video-background store is migrated
@@ -628,14 +629,24 @@ const HOLIDAY_ASSETS_DIR = "dsh-background-by-model/holiday";
 */
 const HOLIDAY_ASSET_HOSTS = [`https://cdn.jsdelivr.net/gh/${HOLIDAY_ASSETS_REPO}@${HOLIDAY_ASSETS_REF}/${HOLIDAY_ASSETS_DIR}`, `https://raw.githubusercontent.com/${HOLIDAY_ASSETS_REPO}/${HOLIDAY_ASSETS_REF}/${HOLIDAY_ASSETS_DIR}`];
 /**
-* Sub-directory of the store holding downloaded festival art.
+* Sub-directory of the store holding downloaded festival art: one directory per
+* release of the asset repository, one file per asset inside it.
 *
 * It is NOT part of the slot namespace and must never become part of it: the
-* cached files are named after the ASSET (`mid-autumn.webp`), not after a slot,
-* so nothing here can ever be mistaken for a user's `modelbg-<slot>`. That is
-* what lets a holiday slot stay read-only — `writeImage` still refuses it, and
-* the store still never holds a `modelbg-h-*` file — while the art itself is
-* now fetched rather than shipped.
+* cached files are named after the ASSET (`mid-autumn.webp`) and their parent
+* after the release the plugin pins (`v0.1.0`), never after a slot, so nothing
+* here can ever be mistaken for a user's `modelbg-<slot>`. That is what lets a
+* holiday slot stay read-only — `writeImage` still refuses it, and the store
+* still never holds a `modelbg-h-*` file — while the art itself is now fetched
+* rather than shipped.
+*
+* The version segment is load-bearing. A cache keyed by asset name alone answers
+* a NEW generation with the OLD bytes and says nothing about it, on a day nobody
+* is watching — which is the failure the pinned reference exists to prevent, one
+* layer further down. With the tag in the path, a build that pins a new tag looks
+* in a directory nobody has written to yet, so the first holiday after an update
+* is a cache MISS instead of a stale hit. What happens to the generation that
+* leaves behind is `pruneHolidayCache`.
 */
 const HOLIDAY_CACHE_DIR = "holiday-cache";
 /** Cap and timeout for one asset download; these files are 228 KB and 500 KB. */
@@ -729,7 +740,7 @@ const dataDir = () => dshHomePath(DATA_DIR);
 const configPath = () => dshHomePath(DATA_DIR, CONFIG_FILE);
 const legacyWallpaperPath = () => dshHomePath(DATA_DIR, LEGACY_WALLPAPER);
 const imagePath = (slot) => dshHomePath(DATA_DIR, `${IMAGE_PREFIX}${slot}`);
-const holidayCachePath = (asset) => dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR, asset);
+const holidayCachePath = (asset) => dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR, HOLIDAY_ASSETS_REF, asset);
 const exists = async (p) => {
 	try {
 		await access(p);
@@ -1223,16 +1234,48 @@ async function fetchPresetImage(slot) {
 async function writeHolidayCache(asset, buf) {
 	try {
 		await ensureDir();
-		await mkdir(dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR), { recursive: true });
+		await mkdir(dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR, HOLIDAY_ASSETS_REF), { recursive: true });
 		const target = holidayCachePath(asset);
 		const temp = `${target}.tmp`;
 		await writeFile(temp, buf);
 		await rename(temp, target);
+		await pruneHolidayCache();
 	} catch (e) {
 		console.warn(`dsh-background-by-model: could not cache the "${asset}" holiday art`, e);
 	}
 }
 /**
+* Drop every generation of the cache except the one this build reads.
+*
+* Called AFTER a successful download, never before it, so a generation is only
+* ever removed once its replacement is on disk: a pinned release that turns out
+* to be unreachable must not cost the user the art they already had.
+*
+* It walks the cache ROOT and removes every entry whose name is not this build's
+* tag, so the generation in use cannot be touched however the comparison goes:
+* the entries it compares are generation names, never asset names. That also
+* makes an upgrade clean up after itself — 0.7.9 cached flat `holiday-cache/<asset>`
+* files, which this build never reads and this sweep removes.
+*
+* Best-effort, like the write that calls it: a cache that could not be pruned is
+* a cache with one extra entry in it, and failing a festival over that would be
+* the tail wagging the dog.
+*/
+async function pruneHolidayCache() {
+	try {
+		const root = dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR);
+		for (const entry of await readdir(root)) {
+			if (entry === HOLIDAY_ASSETS_REF) continue;
+			await rm(dshHomePath(DATA_DIR, HOLIDAY_CACHE_DIR, entry), {
+				recursive: true,
+				force: true
+			});
+		}
+	} catch (e) {
+		console.warn("dsh-background-by-model: could not prune the older holiday art", e);
+	}
+}
+/**
 * The wallpaper a holiday slot resolves to, or null when the slot is not a
 * holiday's.
 *
@@ -1281,10 +1324,11 @@ async function listSlots() {
 * makes the profile's own slots free to write, and it is why this feature needs
 * no slot remapping: there is nothing left to collide with.
 *
-* `holiday-cache/` is deliberately untouched. It is not the user's
-* configuration: it is downloaded festival art that belongs to no slot, costs a
-* network round trip to get back, and has nothing to do with which profile is
-* selected.
+* `holiday-cache/` is deliberately untouched — every generation of it. It is not
+* the user's configuration: it is downloaded festival art that belongs to no
+* slot, costs a network round trip to get back, and has nothing to do with which
+* profile is selected. Pruning superseded generations is `pruneHolidayCache`'s
+* job, and it is not this one.
 *
 * Best-effort per entry — one unremovable file must not abort the rest and leave
 * the store half wiped — but the failures are logged rather than swallowed,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-background-by-model",
-  "version": "0.7.9",
+  "version": "0.7.10",
   "description": "Appearance plugin for DeepSeek Harness: an ordered list of model rules, each carrying its own wallpapers — every image with its own framing and theme color — plus layout mode, opacity and blur; the background follows the model you switch to, and a rule can rotate through several images on its own rhythm.",
   "keywords": [
     "deepseek-harness",