tosijs-ui 1.6.22 → 1.6.23

package/CHANGELOG.md

@@ -1,5 +1,66 @@
 # Changelog
 
+## 1.6.23
+
+**The two biggest sources of ecosystem-wide friction — runaway dev servers and haltija
+conflicts — fixed in the version people actually run.** If you use `tosijs-ui/site`, update.
+
+### Machine-safety guards (a leaking/forgotten dev server took a machine down twice)
+
+- **`killStrayServer` no longer `kill -9`s processes merely _connected_ to its port.** It freed
+  its port with `lsof -ti:PORT | xargs kill -9`, and `lsof -i:PORT` matches sockets whose local
+  **or remote** port is PORT — so it SIGKILLed every _client_ of that port, not just the
+  listener: the browser reading your page, an editor's language server, anything with a
+  connection. It now signals **only the process listening** on the port, only after confirming
+  it is a JS runtime, SIGTERM first and SIGKILL only as a fallback.
+
+- **Machine-health preflight** before each build and at dev-server launch (and on a 60s tick):
+  refuses to add load to a machine that is already dying — a runaway dev server, a VM stall.
+  Names the offending PIDs, sizes, ages and project dirs. **Local LLM runtimes (LM Studio,
+  ollama, llama.cpp, mlx_lm, vLLM, …) are exempt** — their memory shrinks the dev budget rather
+  than tripping the guard, and a big _non-dev_ process (Docker, a JVM, a database) only warns,
+  never fails. It never kills anything; it declines to _start_. `preflight: 'fail' | 'warn' |
+  false` in the site config, or `DEV_SKIP_PREFLIGHT=1`; auto-downgraded to a warning in CI.
+
+- **`idleTimeoutHours` (default 8)** — ⚠️ **behavior change:** the dev server now exits after 8
+  idle hours (no request, no rebuild). A forgotten dev server is not inert — it is a days-old
+  process still running the code it loaded at launch (updating the package does nothing for it).
+  `idleTimeoutHours: 0` disables it.
+
+- **RSS ceiling** (`memoryLimitMb`) sampled every 60s, and a **rebuild-loop detector** that
+  stops a build which writes a file it also watches instead of spawning a bundler forever.
+
+### haltija
+
+- Spawned as a **pinned range (`haltija@^1.4.0`)**, overridable with `HALTIJA_VERSION`, instead
+  of a floating `@latest` fetched fresh on every adopter's machine. The 1.4.0 floor is
+  deliberate: it routes `hj` by working directory (so a project's dev server drives its own
+  browser, not whichever was focused), never overwrites a newer machine-wide `hj`, and exits
+  non-zero on a failed command — which the doc-test lane now checks instead of racing to a
+  timeout. Teardown kills only its own process tree, never a haltija you were already running.
+
+### Also fixed
+
+- **Builds again against tosijs 1.6.9.** `<tosi-example>` detected hydration with a
+  `try { this.parts.js } catch` probe, which before tosijs 1.6.9 permanently poisoned the
+  `parts` proxy on the pre-hydration read it performed (it could blank the whole doc page), and
+  whose `private get hydrated()` collided with 1.6.9's new public `hydrated` (a `tsc` error, so
+  the package could not build against current tosijs). The probe is gone — it uses the base
+  class's `this.hydrated` now — and the tosijs floor moves to **`^1.6.9`**
+  ([tosijs#13](https://github.com/tonioloewald/tosijs/issues/13)).
+
+- **The doc-browser no longer double-suffixes the page `<title>`.** On path routing it re-derived
+  the title on hydration, ignoring `headTitle` and re-appending the project name to a title that
+  already contained it — so the home page flipped from its real title to `tosijs-ui — tosijs-ui`
+  the moment the bundle loaded. The static generator and the runtime now share **one**
+  `pageTitle` rule (case-insensitive, both-directions containment check; whitespace-only titles
+  and project names handled). Fixes [#6](https://github.com/tonioloewald/tosijs-ui/issues/6).
+
+- **Two more per-rebuild build leaks moved to child processes** (the dev server lives for days):
+  the live-example **check** (`new AsyncFunction` per block + a `Bun.Transpiler` per `ts` block)
+  and the bundle **gzip** (native zlib). Both now run in a child whose memory the OS reclaims on
+  exit; the gzip stays zlib-in-a-child (not the `gzip` CLI) so the reported size doesn't drift.
+
 ## 1.6.22
 
 ### Fixed

package/dist/doc-browser.js

@@ -155,6 +155,7 @@ import { elements, vars, varDefault, bindings, touch, getListItem, debounce, tos
 import { buildSlugMap, pathForSlug, filenameForPath } from './doc-system/routing';
 import { buildNavTree } from './doc-system/nav-tree';
 import { renderDocMarkdown } from './doc-system/render';
+import { pageTitle } from './doc-system/doc-title';
 import { LiveExample, testManager } from './live-example';
 import { tosiSidenav, TosiSidenav } from './side-nav';
 import { icons } from './icons';
@@ -613,9 +614,11 @@ export function createDocBrowser(options) {
         LiveExample.insertExamples(docContent, context, doc.path || undefined);
         scrollToHashExample();
         if (routing === 'path') {
-            document.title = projectName
-                ? `${doc.title} — ${projectName}`
-                : doc.title;
+            // The SAME rule the static generator used for this page's <head> (doc-title.ts).
+            // This used to re-derive it — ignoring `headTitle` and re-suffixing a title that
+            // already ended in the project name — so the home page's title flipped to
+            // "tosijs-ui — tosijs-ui" the moment the bundle loaded (issue #6).
+            document.title = pageTitle(doc, projectName);
         }
     };
     // Always resolve to the RAW doc from the original array — docs reached via the

package/dist/doc-system/doc-title.d.ts

@@ -0,0 +1,19 @@
+/** The fields of a Doc that bear on its title. Structural, so both callers fit. */
+export interface TitledDoc {
+    title: string;
+    /** verbatim <title>, no project suffix — a per-doc override */
+    headTitle?: string;
+}
+/**
+ * `headTitle` wins verbatim. Otherwise suffix the project name — but only when the two
+ * are genuinely distinct.
+ *
+ * "Distinct" means **neither string contains the other, compared case-insensitively.**
+ * A one-way, case-sensitive `title.includes(projectName)` check missed real cases: a
+ * project titled `Tosijs-UI` against a doc titled `tosijs-ui` (case), and a doc whose
+ * title is a *substring* of the project name (the containment runs the other way). Any
+ * of those, suffixed, reads as redundant — "tosijs-ui — tosijs-ui" being the one that
+ * shipped (issue #6). When one already contains the other, the doc's own title is the
+ * more specific label, so use it alone.
+ */
+export declare function pageTitle(doc: TitledDoc, projectName?: string): string;

package/dist/doc-system/doc-title.js

@@ -0,0 +1,51 @@
+/*
+The ONE rule for a doc page's <title>.
+
+It has to be one rule, because it is applied twice — once by the static generator into
+the pre-rendered `<head>`, and once by the doc-browser when it hydrates. Two
+independently-written copies drifted, and the title became the one thing on the page
+that visibly MOVED on hydration:
+
+    JS off        →  "tosijs-ui — robust, dependency-free web components"   (headTitle)
+    hydrated      →  "tosijs-ui — tosijs-ui"                                (!)
+
+...because the runtime copy ignored `headTitle` and re-suffixed a title that already
+ended in the project name. That is GitHub issue #6, and it was live on our own home
+page while this release's whole thesis was "hydration is purely additive, nothing
+moves".
+
+Shared by `site/generate-site.ts` (build) and `doc-browser.ts` (runtime). Pure — no DOM,
+no Bun — so both can import it.
+*/
+/**
+ * `headTitle` wins verbatim. Otherwise suffix the project name — but only when the two
+ * are genuinely distinct.
+ *
+ * "Distinct" means **neither string contains the other, compared case-insensitively.**
+ * A one-way, case-sensitive `title.includes(projectName)` check missed real cases: a
+ * project titled `Tosijs-UI` against a doc titled `tosijs-ui` (case), and a doc whose
+ * title is a *substring* of the project name (the containment runs the other way). Any
+ * of those, suffixed, reads as redundant — "tosijs-ui — tosijs-ui" being the one that
+ * shipped (issue #6). When one already contains the other, the doc's own title is the
+ * more specific label, so use it alone.
+ */
+export function pageTitle(doc, projectName) {
+    // A whitespace-only `headTitle` is not an override, it's a blank <title> — trim it
+    // and fall through. Trimming also cleans up an accidentally-padded one ("verbatim"
+    // means "no project suffix", not "keep your stray whitespace").
+    const headTitle = doc.headTitle?.trim();
+    if (headTitle)
+        return headTitle;
+    // Trim first: a whitespace-only name ('' or '   ') is "no project name", not a real
+    // one. Untrimmed, ' ' is truthy and slips through — and worse, `title.includes(' ')`
+    // is true for any multi-word title, so it would silently suppress a suffix that
+    // should appear. Trimming also strips a padded 'tosijs-ui ' down to the real name.
+    const project = projectName?.trim() ?? '';
+    if (!project)
+        return doc.title;
+    const t = doc.title.toLowerCase();
+    const p = project.toLowerCase();
+    if (t.includes(p) || p.includes(t))
+        return doc.title;
+    return `${doc.title} — ${project}`;
+}

package/dist/doc-system/site/check-examples-cli.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/doc-system/site/check-examples-cli.js

@@ -0,0 +1,39 @@
+/*
+Live-example build check — the subprocess entry point for checkExamples().
+
+Why a child process, when this is "just" compiling some strings:
+
+checkExamples() compiles EVERY executable block in the corpus (83 of them here) with
+`new AsyncFunction(js)` on EVERY dev rebuild, and — for `ts` blocks — runs them
+through a `Bun.Transpiler`. Both retain memory the JS heap never shows you:
+
+  - JSC caches compiled code, KEYED BY SOURCE TEXT. That last part is the trap: replay
+    the same corpus and RSS plateaus (the cache dedups), so a leak test on an unchanged
+    corpus reports "flat" and lies. But a dev server only rebuilds BECAUSE A FILE
+    CHANGED — the source is fresh every time, so nothing dedups and it never plateaus.
+    Measured with fresh sources: +7.1MB over 40 rebuilds, still climbing.
+  - `Bun.Transpiler` strands ~40KB per construction (oven-sh/bun#34053).
+
+In a process that lives for days across thousands of rebuilds, that is unbounded. The
+child hands all of it back to the OS on exit. Same reasoning — and the same fix — as
+the bundle step, the ePub step, and generate-css; see the Bun.build note in
+orchestrator.ts.
+
+  bun check-examples-cli.ts <docs.json> [contextKey ...]
+  → writes ExampleProblem[] as JSON to stdout; exit 0 whether or not problems were
+    found (a problem is data, not a crash — only a real failure exits non-zero).
+
+Build-time only. Never import this from browser code.
+*/
+import { checkExamples } from './check-examples';
+const docsJson = process.argv[2];
+if (!docsJson) {
+    console.error('check-examples-cli: expected a docs.json path');
+    process.exit(1);
+}
+const contextKeys = process.argv.slice(3);
+const corpus = JSON.parse(await Bun.file(docsJson).text());
+const problems = await checkExamples(corpus, contextKeys.length ? { contextKeys } : {});
+// stdout is the channel — the parent parses this. Anything else this process prints
+// (warnings from the transform, say) goes to stderr so it can't corrupt the payload.
+process.stdout.write(JSON.stringify(problems));

package/dist/doc-system/site/dev-server.d.ts

@@ -2,6 +2,32 @@ import type { SiteConfig } from './site-config';
 declare global {
     var Bun: any;
 }
+/**
+ * Resolve the idle-exit timeout to milliseconds (0 = disabled).
+ *
+ * Env wins over config, config over the default. An unparseable value falls back
+ * to the default rather than to 0: a typo'd `DEV_IDLE_TIMEOUT_HOURS=8h` must not
+ * silently turn the guard OFF — that is the exact failure it exists to prevent.
+ * Only an explicit non-positive number disables it.
+ */
+export declare function resolveIdleMs(configHours: number | undefined, envHours: string | undefined): number;
+/**
+ * Resolve the RSS ceiling in MB (0 = disabled).
+ *
+ * Same rule as `resolveIdleMs`, and it was NOT being applied: the ceiling was read as
+ * `Number(env ?? config ?? 4096)`, which fails in both directions at once —
+ *
+ *   DEV_MEMORY_LIMIT_MB=''   → `??` passes '' through → Number('') === 0 → the ceiling
+ *                              is ZERO, so `rss >= limit` is true on the first sample
+ *                              and the dev server kills itself on every rebuild.
+ *   DEV_MEMORY_LIMIT_MB=4gb  → NaN → every `>=` comparison is false → the guard is
+ *                              silently OFF, on the machine of someone who was
+ *                              explicitly trying to configure it.
+ *
+ * An empty env var is *unset*, not zero. Garbage falls back to the default, never to
+ * off. Only an explicit non-positive number disables the ceiling.
+ */
+export declare function resolveLimitMb(configMb: number | undefined, envMb: string | undefined): number;
 export declare function devServer(config: SiteConfig, opts?: {
     test?: boolean;
     build?: () => unknown | Promise<unknown>;