diff --git a/AGENTS.md b/AGENTS.md
index ca9fa0ba17..1ed4d681ba 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -5,7 +5,7 @@
 - Describe what Marko does, never what it lacks. No "there is no X", and no workarounds that route around the framework.
 - Do not document compile errors. The compiler reports them with a code frame. This covers content that reframes them, such as a table mapping another framework's attribute names onto Marko's.
 - Do not document errors, warnings, or checks that only a `MARKO_DEBUG` build produces. State the rule, not the message.
-- Never document a known bug as intended behavior. Grep `agent-feedback/bugs.md` in the `marko` repo first; a hit means file the defect and leave the docs alone.
+- Never document a known bug as intended behavior. Grep `agent-feedback/` in the `marko` repo first; a hit means file the defect and leave the docs alone.
 - Drop a topic that loses its main claim to the rule above. What remains is the least verified part of it.
 - Prefer filing a defect to writing a caveat that teaches readers to live with one.
 
@@ -121,4 +121,4 @@
 
 ## Agent feedback
 
-Anything actionable but out of scope for the current task — a suspected bug, cleanup, a perf/size win, tooling friction, or code that was confusing — must be recorded in [`agent-feedback/`](agent-feedback/README.md) before finishing. Don't silently drop it, and don't fix it inside an unrelated diff.
+Anything actionable but out of scope for the current task (suspected bug, cleanup, perf or size win, tooling friction, confusing code) must be filed in [`agent-feedback/`](agent-feedback/README.md) before finishing. Never drop it silently. Never fix it inside an unrelated diff.
diff --git a/agent-feedback/README.md b/agent-feedback/README.md
index c79150a904..53b82b15a7 100644
--- a/agent-feedback/README.md
+++ b/agent-feedback/README.md
@@ -1,39 +1,65 @@
 # Agent Feedback
 
-Actionable observations that were **out of scope for the task that surfaced them**. If something is in scope, fix it instead. Do not expand a task's diff to fix issues recorded here.
+Actionable observations that were out of scope for the task that surfaced them. In scope: fix it. Out of scope: file it here. Never expand a task's diff to fix an item recorded here.
 
-## When to add an entry
+One item per file in `items/`, named `YYYY-MM-DD-<slug>.md`.
 
-While working on any task, record anything a future contributor should act on:
+## When to file
 
-- a suspected bug you couldn't pursue → `bugs.md`
-- duplication, dead code, inconsistency, refactor opportunities → `cleanup.md`
-- runtime speed or bundle size opportunities → `perf.md`
-- friction in builds, tests, tooling, or repo workflows → `dx.md`
-- code or docs that were confusing, and what would have clarified them → `unclear.md`
+Anything a future contributor should act on:
 
-## Rules
+- `bug`: a suspected defect left unpursued
+- `cleanup`: duplication, dead code, inconsistency, refactor opportunity
+- `perf`: speed, memory, payload or bundle size, build time
+- `dx`: friction in builds, tests, tooling, or repo workflows
+- `unclear`: code or docs that were confusing, and what would have clarified them
 
-1. **Search the category file first.** If an entry already covers it, don't duplicate; append a corroborating sentence only if you have new information.
-2. **Be self-contained.** Include enough detail (paths, symbols, reasoning) that someone can act without re-discovering your analysis. Never reference "my earlier analysis" or conversation context.
-3. **Cite by stable symbol, not line number.** Line numbers rot with the next edit; anchor the primary citation to the nearest enclosing stable symbol (exported function, class, variable, or a heading for docs). A line number may appear in the body as a secondary hint.
-4. **Append to the end** of the category file.
-5. Entries are **removed when resolved** (delete, don't mark done; git history is the archive).
-6. **Verify before recording.** A guess is not feedback.
+## Rules
 
-## Resolving a "won't fix" item
+1. **Verify first.** A guess is not feedback. Every item ends with a check that reproduces the claim.
+2. **Dedupe first.** `grep -ril '<path or symbol>' agent-feedback/items`. If a file covers it, edit that file only when you add new information.
+3. **Check the code site.** An intent comment there means the behavior is deliberate. Do not file it.
+4. **Self-contained.** Paths, symbols, reasoning. Never reference conversation context or "earlier analysis".
+5. **Cite by stable symbol**, never line number.
+6. **State the defect and the check.** Never describe what works. Never narrate a landed fix.
+7. **Direction is preventive for `unclear` and `dx`.** Name what would have stopped the trip: a comment, a doc line, a lint rule, a compile error, a debug-only warning. The goal is that the next agent does not hit it.
+8. **Resolve by deleting the file in the same PR as the fix.** A partial fix rewrites the file to what remains.
+9. **Won't-fix is a maintainer's call, never an agent's.** Add a comment (two lines max) at the code site stating the behavior and why it is deliberate, then delete the file. The comment is what stops re-filing. Never consult git history to learn whether something was resolved; if it is not in `items/` and not commented at the site, it is unresolved.
 
-When a maintainer has explicitly deemed an item "won't fix" / "not worth it", resolve it by adding a brief inline comment at the code site that captures the decision (so it is not re-filed), then remove the entry. Only on such an explicit call — never on your own initiative.
+## Item format
 
-## Entry format
+`items/YYYY-MM-DD-<slug>.md`:
 
 ```md
-## <one-line imperative summary>
+---
+type: bug | cleanup | perf | dx | unclear
+impact: high | med | low
+effort: high | med | low
+site: <path/to/file.ts> › <nearestStableSymbol>
+---
+
+# <one-line imperative title>
 
-`<primary/file/path.ts>` › `<nearestStableSymbol>` | 2026-07-02 | impact:<low|med|high> | effort:<low|med|high>
+<2-6 sentences: the problem, why it matters, a concrete direction. Cut evidence a fixer can re-derive from the site.>
 
-<2–6 sentences: the problem, why it matters, and a concrete suggested direction,
-ending with the check that re-verifies the claim (a command, input, or
-observation). Cut evidence beyond what a fixer needs to act; further detail is
-re-derived from the citation. Additional file paths inline as needed.>
+Check: <command, input, or observation that reproduces the claim>
 ```
+
+`impact`: what breaks or is lost if ignored. `effort`: expected size of the fix. Both are the filer's estimate; triage re-judges.
+
+## Repo notes
+
+The markojs.com site: docs as markdown under `docs/`, a `@marko/run` app under `src/`, plus the browser playground. The root `AGENTS.md` holds the documentation writing guidelines and they govern any docs change; read it too.
+
+**Reproduce a claim.**
+
+- Docs content: `grep` the page, then verify the behavior it describes against Marko runtime source, a fixture snapshot, or a compiled probe. An existing docs sentence is not evidence.
+- Anchors and slugs: `github-slugger` resolves collisions in document order, so a second heading with the same slug becomes `#name-1`. Check with the slugger directly rather than guessing.
+- Playground and site behavior: `pnpm run dev`, then reproduce in the browser.
+- Build-derived output (page `<meta>`, generated llms bundles): `pnpm run build`, then read `src/routes/docs/_compiled-docs/`.
+
+**Guard tests.** `pnpm test` (`vitest run`). A docs-structure claim (every page indexed, every anchor resolving) is best guarded by a test that walks `docs/**/*.md`, not by a snapshot of one page.
+
+**Pre-ship.** `pnpm run lint` (markdownlint over `docs/**/*.md`, then cspell) and `pnpm run type-check` (`marko-type-check`). `pnpm run build` when the change can affect generated output.
+
+**Gotchas.** `pnpm run lint` runs cspell, which has a Node engine floor above some default installs and aborts on the version check before checking anything; the husky pre-commit hook hits the same wall. Never document a known Marko bug as intended behavior: file the defect instead. Never use an em dash in docs, and grep `' - '` too, since a spaced hyphen doing the same job is the same problem.
diff --git a/agent-feedback/bugs.md b/agent-feedback/bugs.md
deleted file mode 100644
index 08a3b8c577..0000000000
--- a/agent-feedback/bugs.md
+++ /dev/null
@@ -1,15 +0,0 @@
-# Suspected Bugs
-
-Out-of-scope defects noticed while working on something else. Format and rules: [README.md](README.md).
-
-## Point `core-tag.md`'s `<style>` links at the tag section instead of the `style=` attribute section
-
-`docs/reference/core-tag.md` › `## <html-script> & <html-style>` | 2026-07-29 | impact:low | effort:low
-
-`docs/reference/native-tag.md` has two headings that slug to the same base: the `style=` heading under Enhanced Attributes and the `<style>` heading under Enhanced Tags. `github-slugger` resolves collisions in document order, so the attribute section (line 56) takes `#style` while the tag section (line 533) becomes `#style-1`. Both `./native-tag.md#style` links in `core-tag.md` mean the `<style>` tag but land the reader on the unrelated `style=` attribute. `#script` is unaffected, since the `<script>` heading is the only one slugging to it. Prefer renaming one heading over hardcoding `#style-1`, which would rot silently if heading order ever changed. Re-verify from the repo root: `node --input-type=module -e "import S from './node_modules/github-slugger/index.js'; const s = new S(); for (const h of ['class=','style=','Event Handlers','<script>','<style>']) console.log(h, '->', s.slug(h))"` prints `style= -> style` and `<style> -> style-1`.
-
-## Add `NaN` and `0n` to the values that drop a `style=` declaration
-
-`docs/reference/native-tag.md` › `### style=` | 2026-07-30 | impact:low | effort:low
-
-The `style=` section states "A declaration is dropped when its value is `false`, `null`, `undefined` or an empty string, but `0` is kept" (line 90), which reads as the complete set but omits `NaN` and bigint `0n`. Marko's runtime writes a declaration only when `value || value === 0` (`stringifyStyleObject` in the marko repo's `packages/runtime-tags/src/common/helpers.ts`, where a bigint never satisfies `0n === 0`), so `<div style={ "line-height": total / count }/>` with a `count` of `0` renders no `line-height` declaration at all rather than the ignored `line-height:NaN` the current wording promises, and a `0n` value disappears the same way while the equivalent attribute still writes `NaN` or `0`. `docs/reference/language.md` already lists all six values under [Skipped Values](./language.md#skipped-values) for dynamic text and contrasts them with skipped attributes, so the `style=` sentence should carry the same set and link there; the neighboring `class=` claim ("every falsy value drops the class") needs no change, since a class name is kept only when its value is truthy. Re-verify: `grep -n "A declaration is dropped" docs/reference/native-tag.md` lists four values against the six in `grep -n "renders nothing" docs/reference/language.md`, then run `pnpm run lint`.
diff --git a/agent-feedback/cleanup.md b/agent-feedback/cleanup.md
deleted file mode 100644
index 24bb0d0d7a..0000000000
--- a/agent-feedback/cleanup.md
+++ /dev/null
@@ -1,42 +0,0 @@
-# Cleanup
-
-Duplication, dead code, inconsistencies, refactor opportunities. Format and rules: [README.md](README.md).
-
-## Drop the deprecated `@types/flexsearch` stub
-
-`package.json` › `devDependencies` | 2026-07-27 | impact:low | effort:low
-
-`@types/flexsearch` is a deprecated stub package, and pnpm prints a deprecation
-warning on every install. `flexsearch` has shipped its own types since 0.8: its
-`package.json` declares both `types` and `exports["."].types` as `./index.d.ts`.
-The only consumer is `src/util/search-worker.ts`, which imports
-`flexsearch, { type Index }`. Removing the dependency was verified to leave
-`marko-type-check` reporting no errors in `search-worker.ts`. Re-verify with
-`pnpm remove @types/flexsearch && pnpm exec marko-type-check`.
-
-## Fill in the three docs that have no usable opening prose
-
-`docs/explanation/optimizing-performance.md` › `# Optimizing Performance` | 2026-07-28 | impact:med | effort:med
-
-`optimizing-performance.md` and `docs/guide/publishing-components.md` are
-heading-only stubs with no body text at all, and `docs/guide/library-integration.md`
-opens its first section with the bare fragment "Use `<lifecycle>` tag" instead of
-a sentence. Beyond the reader-facing gap, the build derives each page's
-`<meta name=description>` from its opening paragraph, so the two stubs fall back
-to the generic site description and library-integration advertises itself to
-search engines as "Use <lifecycle> tag". Writing an intro paragraph for each
-fixes both at once. Re-verify with `npm run build` and then
-`grep -h '"description"' src/routes/docs/_compiled-docs/{explanation/optimizing-performance,guide/publishing-components,guide/library-integration}+meta.json`.
-
-## Remove the empty token-type branch in the markodown walker
-
-`src/util/markodown.ts` › `markoDocs` | 2026-07-29 | impact:low | effort:low
-
-Inside `markoDocs()`'s `walkTokens`, after the code-fence handling there is a
-chain that begins `if (token.type === "code" && token.lang === "marko") {}`
-with an empty body, followed by an `else if (token.type === "codespan")`
-branch whose only content is a commented-out line. Only the final
-`else if (token.type === "link")` branch does anything. The two dead branches
-read as if they carry behavior and slow down anyone tracing how tokens are
-transformed; collapsing the chain to a single `if (token.type === "link")`
-preserves behavior. Re-verify by reading the chain and running `pnpm test`.
diff --git a/agent-feedback/dx.md b/agent-feedback/dx.md
deleted file mode 100644
index e8b7964c77..0000000000
--- a/agent-feedback/dx.md
+++ /dev/null
@@ -1,33 +0,0 @@
-# Developer Experience
-
-Friction in builds, tests, tooling, or repo workflows. Format and rules: [README.md](README.md).
-
-## Map playground runtime errors back to the user's source
-
-`src/util/workspace.ts` › `onRuntimeError` | 2026-07-27 | impact:med | effort:med
-
-When an `ErrorEvent` carries no `error`, the handler synthesizes an `Error` whose
-message appends `ev.filename`, `ev.lineno` and `ev.colno`. Those coordinates
-point into the generated bundle, so the playground reports a location such as
-`server.js:1,38842`, which is real but names no file the user can open. Build
-errors avoid this because marko's `CompileError` frame already names the authored
-file, and the error output renders that frame. The workspace keeps sourcemaps for
-both the preview modules and the compiled templates, so the position could be
-mapped back to the authored file and line before the error is stored, letting
-runtime errors render the same code frame as build errors. Re-verify by entering
-a template that throws on render, such as calling a string method on a number,
-and reading the location shown in the runtime error card.
-
-## Root the playground filesystem at `/` instead of `/app/tags`
-
-`src/util/workspace.ts` › `rootDir` | 2026-07-28 | impact:med | effort:med
-
-The virtual filesystem hangs every user file off `` const rootDir = `${projectDir}/tags/` `` (`projectDir` is `/app`), so paths surface to the user as `app/tags/index.marko` — in the file tabs, and inside every compile error's `at …` location. Neither segment means anything to someone writing a template in the playground: `app` exists only to give `package.json` a home, and `tags/` exists only because the compiler discovers custom tags in a directory named `tags`. Conceptually the files live at the project root, and the compiler can be told so directly by passing `tagDiscoveryDirs` (or a taglib `taglib-imports`/`tags-dir` entry) that names `/` as a tag directory, which is the same mechanism `@marko/compiler`'s taglib finder already uses for real projects. That would let paths render as `index.marko`.
-
-The blocker is that share links encode file paths: `src/util/hasher.ts` compresses the file list into the URL hash, so shortening the root changes what every existing `#…` link decodes to. Any change needs either a migration that rewrites legacy paths on decode (strip a leading `app/tags/`) or a hash version bump — `hasher.ts` already distinguishes a `v2` payload, so the versioning hook exists. Re-verify the current behavior by opening the playground and reading the path in a compile error's `at` line, or the tab titles, which today read `app/tags/…`.
-
-## Pre-commit lint fails on machines whose default Node is below cspell's floor
-
-`package.json` › `scripts.lint` | 2026-08-18 | impact:med | effort:low
-
-`pnpm run lint` (and therefore the husky pre-commit hook) dies with "Unsupported NodeJS version (22.14.0); >=22.18.0 is required" before cspell checks anything when the `node` that pnpm spawns is older than cspell's engine floor, even when a newer Node is first on `PATH` (the mise-shimmed pnpm resolves its own). The repo declares no `engines`/`.nvmrc`/mise config pinning Node for contributors, so nothing surfaces the requirement until the hook blocks a commit. Declaring the required Node version in the repo (`engines` + a mise/`.nvmrc` pin) or running cspell via a version-agnostic entry would make the gate reproducible. Verified 2026-08-18: markdownlint passes, cspell aborts on the version check under Node 22.14, passes under 24.19.
diff --git a/agent-feedback/items/2026-07-19-llms-txt-completeness.md b/agent-feedback/items/2026-07-19-llms-txt-completeness.md
new file mode 100644
index 0000000000..17b02de7fa
--- /dev/null
+++ b/agent-feedback/items/2026-07-19-llms-txt-completeness.md
@@ -0,0 +1,12 @@
+---
+type: unclear
+impact: med
+effort: low
+site: public/llms.txt › Reference
+---
+
+# Index every public docs page in `llms.txt` and add a completeness check
+
+`public/llms.txt` is a hand-maintained static index with no generator, and three substantive on-disk pages are linked from none of its topic sections: `docs/reference/lazy-loading.md` (the whole `import ... with { load: "visible#hero" }` client code-splitting and deferred-hydration API), `docs/explanation/class-vs-tags-api.md` (the flagship Marko 5-to-6 migration mental-model guide), and `docs/reference/supported-environments.md`. The only aggregate generator, `src/routes/docs/_llms/reference-full%2emd+handler.ts`, hardcodes a 9-page list that includes lazy-loading but omits supported-environments, so even the bundle llms.txt labels "Complete reference documentation" is itself incomplete, and class-vs-tags-api appears in no aggregate at all. Because llms.txt is an agent's primary topic index, an agent asked to add deferred client hydration scans the Reference links, finds no lazy-loading entry, and concludes the capability does not exist; an agent porting a Marko 5 class app never discovers class-vs-tags-api.md and mis-migrates from memory. Add the three pages to their sections, add supported-environments to the reference-full list, and add a test asserting every `docs/**/*.md` slug appears in llms.txt so new pages cannot silently drop out.
+
+Check: `grep -c lazy-loading public/llms.txt` is 0 while `docs/reference/lazy-loading.md` exists; same for `class-vs-tags-api` and `supported-environments`.
diff --git a/agent-feedback/items/2026-07-19-template-mount-node-required.md b/agent-feedback/items/2026-07-19-template-mount-node-required.md
new file mode 100644
index 0000000000..11945409de
--- /dev/null
+++ b/agent-feedback/items/2026-07-19-template-mount-node-required.md
@@ -0,0 +1,12 @@
+---
+type: unclear
+impact: low
+effort: low
+site: docs/reference/template.md › Template.mount(input, node, position?)
+---
+
+# Mark `Template.mount`'s `node` param as required, not `Default: undefined`
+
+The `Template.mount(input, node, position?)` parameter table lists a Default of `{}` for `input` and `"beforeend"` for `position`, both real defaults, but `undefined` for `node`, which reads as "node is optional and defaults to undefined". `node` is required: `packages/runtime-tags/index.d.ts` in the marko repo declares `mount(input, reference: Node, position?)` with `reference` non-optional, and `packages/runtime-tags/src/dom/template.ts` › `mount` defaults only `input = {}` while `reference` has no default and is dereferenced unconditionally, so `template.mount(input)` is a type error and throws at runtime. The doc's own examples always pass a node. Agents treat parameter tables as the machine-readable contract: an agent building a client mount or test harness reads "node: Default undefined", writes `template.mount(input)` expecting a detached mount, and burns a debug cycle, or ships code that crashes on the undefined reference. Replace the `undefined` cell with "required", or split required params from defaulted ones, so the table agrees with the `reference: Node` signature.
+
+Check: read the parameter table under the `Template.mount` heading in `docs/reference/template.md`.
diff --git a/agent-feedback/items/2026-07-27-drop-types-flexsearch.md b/agent-feedback/items/2026-07-27-drop-types-flexsearch.md
new file mode 100644
index 0000000000..e7e244ed25
--- /dev/null
+++ b/agent-feedback/items/2026-07-27-drop-types-flexsearch.md
@@ -0,0 +1,12 @@
+---
+type: cleanup
+impact: low
+effort: low
+site: package.json › devDependencies
+---
+
+# Drop the deprecated `@types/flexsearch` stub
+
+`@types/flexsearch` is a deprecated stub package, and pnpm prints a deprecation warning on every install. `flexsearch` has shipped its own types since 0.8: its `package.json` declares both `types` and `exports["."].types` as `./index.d.ts`. The only consumer is `src/util/search-worker.ts`, which imports `flexsearch, { type Index }`. Removing the dependency was verified to leave `marko-type-check` reporting no errors in `search-worker.ts`.
+
+Check: `pnpm remove @types/flexsearch && pnpm exec marko-type-check`.
diff --git a/agent-feedback/items/2026-07-27-playground-runtime-error-source-mapping.md b/agent-feedback/items/2026-07-27-playground-runtime-error-source-mapping.md
new file mode 100644
index 0000000000..799be2dee9
--- /dev/null
+++ b/agent-feedback/items/2026-07-27-playground-runtime-error-source-mapping.md
@@ -0,0 +1,12 @@
+---
+type: dx
+impact: med
+effort: med
+site: src/util/workspace.ts › onRuntimeError
+---
+
+# Map playground runtime errors back to the user's source
+
+When an `ErrorEvent` carries no `error`, the handler synthesizes an `Error` whose message appends `ev.filename`, `ev.lineno`, and `ev.colno`. Those coordinates point into the generated bundle, so the playground reports a location such as `server.js:1,38842`, which is real but names no file the user can open. Build errors avoid this because marko's `CompileError` frame already names the authored file, and the error output renders that frame. The workspace keeps sourcemaps for both the preview modules and the compiled templates, so the position could be mapped back to the authored file and line before the error is stored, letting runtime errors render the same code frame as build errors.
+
+Check: in the playground, enter a template that throws on render, such as calling a string method on a number, and read the location shown in the runtime error card.
diff --git a/agent-feedback/items/2026-07-28-playground-root-dir.md b/agent-feedback/items/2026-07-28-playground-root-dir.md
new file mode 100644
index 0000000000..a257dbecde
--- /dev/null
+++ b/agent-feedback/items/2026-07-28-playground-root-dir.md
@@ -0,0 +1,12 @@
+---
+type: dx
+impact: med
+effort: med
+site: src/util/workspace.ts › rootDir
+---
+
+# Root the playground filesystem at `/` instead of `/app/tags`
+
+The virtual filesystem hangs every user file off `` const rootDir = `${projectDir}/tags/` `` (`projectDir` is `/app`), so paths surface to the user as `app/tags/index.marko`, in the file tabs and inside every compile error's `at ...` location. Neither segment means anything to someone writing a template in the playground: `app` exists only to give `package.json` a home, and `tags/` exists only because the compiler discovers custom tags in a directory named `tags`. Conceptually the files live at the project root, and the compiler can be told so by passing `tagDiscoveryDirs` (or a taglib `taglib-imports`/`tags-dir` entry) naming `/` as a tag directory, the same mechanism `@marko/compiler`'s taglib finder already uses for real projects. Paths would then render as `index.marko`. The blocker is that share links encode file paths: `src/util/hasher.ts` compresses the file list into the URL hash, so shortening the root changes what every existing `#...` link decodes to. Any change needs either a migration that strips a leading `app/tags/` on decode or a hash version bump; `hasher.ts` already distinguishes a `v2` payload, so the versioning hook exists.
+
+Check: open the playground and read the path in a compile error's `at` line, or the tab titles; both read `app/tags/...` today.
diff --git a/agent-feedback/items/2026-07-28-stub-docs-missing-intro-prose.md b/agent-feedback/items/2026-07-28-stub-docs-missing-intro-prose.md
new file mode 100644
index 0000000000..7e51ee8d15
--- /dev/null
+++ b/agent-feedback/items/2026-07-28-stub-docs-missing-intro-prose.md
@@ -0,0 +1,12 @@
+---
+type: cleanup
+impact: med
+effort: med
+site: docs/explanation/optimizing-performance.md › # Optimizing Performance
+---
+
+# Fill in the three docs that have no usable opening prose
+
+`optimizing-performance.md` and `docs/guide/publishing-components.md` are heading-only stubs with no body text at all, and `docs/guide/library-integration.md` opens its first section with the bare fragment "Use `<lifecycle>` tag" instead of a sentence. Beyond the reader-facing gap, the build derives each page's `<meta name=description>` from its opening paragraph, so the two stubs fall back to the generic site description and library-integration advertises itself to search engines as "Use <lifecycle> tag". Writing an intro paragraph for each fixes both at once.
+
+Check: `pnpm run build`, then `grep -h '"description"' src/routes/docs/_compiled-docs/{explanation/optimizing-performance,guide/publishing-components,guide/library-integration}+meta.json`.
diff --git a/agent-feedback/items/2026-07-29-core-tag-style-anchor.md b/agent-feedback/items/2026-07-29-core-tag-style-anchor.md
new file mode 100644
index 0000000000..887af486f9
--- /dev/null
+++ b/agent-feedback/items/2026-07-29-core-tag-style-anchor.md
@@ -0,0 +1,12 @@
+---
+type: bug
+impact: low
+effort: low
+site: docs/reference/core-tag.md › ## <html-script> & <html-style>
+---
+
+# Point `core-tag.md`'s `<style>` links at the tag section instead of the `style=` attribute section
+
+`docs/reference/native-tag.md` has two headings that slug to the same base: the `style=` heading under Enhanced Attributes and the `<style>` heading under Enhanced Tags. `github-slugger` resolves collisions in document order, so the attribute section takes `#style` while the tag section becomes `#style-1`. Both `./native-tag.md#style` links in `core-tag.md` mean the `<style>` tag but land the reader on the unrelated `style=` attribute. `#script` is unaffected, since the `<script>` heading is the only one slugging to it. Prefer renaming one heading over hardcoding `#style-1`, which would rot silently if heading order ever changed.
+
+Check: from the repo root, `node --input-type=module -e "import S from './node_modules/github-slugger/index.js'; const s = new S(); for (const h of ['class=','style=','Event Handlers','<script>','<style>']) console.log(h, '->', s.slug(h))"` prints `style= -> style` and `<style> -> style-1`.
diff --git a/agent-feedback/items/2026-07-29-markodown-dead-branches.md b/agent-feedback/items/2026-07-29-markodown-dead-branches.md
new file mode 100644
index 0000000000..1c3de0856c
--- /dev/null
+++ b/agent-feedback/items/2026-07-29-markodown-dead-branches.md
@@ -0,0 +1,12 @@
+---
+type: cleanup
+impact: low
+effort: low
+site: src/util/markodown.ts › markoDocs
+---
+
+# Remove the empty token-type branch in the markodown walker
+
+Inside `markoDocs()`'s `walkTokens`, after the code-fence handling there is a chain that begins `if (token.type === "code" && token.lang === "marko") {}` with an empty body, followed by an `else if (token.type === "codespan")` branch whose only content is a commented-out line. Only the final `else if (token.type === "link")` branch does anything. The two dead branches read as if they carry behavior and slow down anyone tracing how tokens are transformed; collapsing the chain to a single `if (token.type === "link")` preserves behavior.
+
+Check: read the chain in `markoDocs`, then run `pnpm test`.
diff --git a/agent-feedback/items/2026-07-30-script-tag-run-order.md b/agent-feedback/items/2026-07-30-script-tag-run-order.md
new file mode 100644
index 0000000000..5c5f039cb2
--- /dev/null
+++ b/agent-feedback/items/2026-07-30-script-tag-run-order.md
@@ -0,0 +1,12 @@
+---
+type: unclear
+impact: med
+effort: low
+site: docs/reference/core-tag.md › ## <script>
+---
+
+# Document when a `<script>` that reads a tag variable first runs
+
+The `<script>` section says the body "is executed first when the template has finished rendering and is mounted in the browser" and re-runs when a referenced tag variable changes, which leaves source order as the only ordering a reader can infer between two `<script>` tags. Marko compiles a `<script>` that reads a tag variable into that variable's own setup, so it first runs when the variable initializes, ahead of an earlier-in-source `<script>` that reads nothing. Given `<let/board=DEFAULT/>`, a first script that only assigns `board` from `localStorage`, and a second that reads `board` and writes it back, the second runs first with `DEFAULT` and overwrites the saved value before the restore reads it: the compiled dom output puts the persist effect inside `_let(...)` while the restore stays a separate `$setup__script`. The page's own `<script=remember/>` example already persists `input.collapsed` to `sessionStorage` without saying when it first fires, so the rule belongs there, stated positively: reading a tag variable ties the effect to that variable's initialization. The paired restore-and-persist idiom belongs under `## <lifecycle>`, whose `onMount`/`onUpdate` handlers already model "set up once, then follow the value".
+
+Check: `grep -n "finished rendering" docs/reference/core-tag.md` is the page's only ordering statement and `grep -rn "localStorage" docs/` returns nothing, so neither the rule nor the idiom is documented; the clobber reproduces in the local playground (`pnpm run dev`).
diff --git a/agent-feedback/items/2026-07-30-style-skipped-values.md b/agent-feedback/items/2026-07-30-style-skipped-values.md
new file mode 100644
index 0000000000..066f014c6d
--- /dev/null
+++ b/agent-feedback/items/2026-07-30-style-skipped-values.md
@@ -0,0 +1,12 @@
+---
+type: bug
+impact: low
+effort: low
+site: docs/reference/native-tag.md › ### style=
+---
+
+# Add `NaN` and `0n` to the values that drop a `style=` declaration
+
+The `style=` section states that a declaration is dropped when its value is `false`, `null`, `undefined`, or an empty string, but `0` is kept. That reads as the complete set and omits `NaN` and bigint `0n`. Marko's runtime writes a declaration only when `value || value === 0` (`stringifyStyleObject` in the marko repo's `packages/runtime-tags/src/common/helpers.ts`, where a bigint never satisfies `0n === 0`), so `<div style={ "line-height": total / count }/>` with a `count` of `0` renders no `line-height` declaration at all rather than the ignored `line-height:NaN` the current wording promises, and a `0n` value disappears the same way while the equivalent attribute still writes `NaN` or `0`. `docs/reference/language.md` already lists all six values under Skipped Values for dynamic text and contrasts them with skipped attributes, so the `style=` sentence should carry the same set and link there. The neighboring `class=` claim needs no change, since a class name is kept only when its value is truthy.
+
+Check: `grep -n "A declaration is dropped" docs/reference/native-tag.md` lists four values against the six in `grep -n "renders nothing" docs/reference/language.md`.
diff --git a/agent-feedback/items/2026-08-18-cspell-node-version-floor.md b/agent-feedback/items/2026-08-18-cspell-node-version-floor.md
new file mode 100644
index 0000000000..4491527987
--- /dev/null
+++ b/agent-feedback/items/2026-08-18-cspell-node-version-floor.md
@@ -0,0 +1,12 @@
+---
+type: dx
+impact: med
+effort: low
+site: package.json › scripts.lint
+---
+
+# Pin the Node version that `pnpm run lint` needs so the pre-commit hook is reproducible
+
+`pnpm run lint`, and therefore the husky pre-commit hook, dies with "Unsupported NodeJS version (22.14.0); >=22.18.0 is required" before cspell checks anything when the `node` that pnpm spawns is older than cspell's engine floor, even when a newer Node is first on `PATH` (the mise-shimmed pnpm resolves its own). The repo declares no `engines`, `.nvmrc`, or mise config pinning Node for contributors, so nothing surfaces the requirement until the hook blocks a commit. Declaring the required Node version in the repo (`engines` plus a mise or `.nvmrc` pin) or running cspell via a version-agnostic entry would make the gate reproducible.
+
+Check: `pnpm run lint` under Node 22.14 aborts on cspell's version check; markdownlint passes and the same command passes under Node 24.19.
diff --git a/agent-feedback/perf.md b/agent-feedback/perf.md
deleted file mode 100644
index 6afb16b982..0000000000
--- a/agent-feedback/perf.md
+++ /dev/null
@@ -1,3 +0,0 @@
-# Performance
-
-Runtime speed and bundle size opportunities. Format and rules: [README.md](README.md).
diff --git a/agent-feedback/unclear.md b/agent-feedback/unclear.md
deleted file mode 100644
index a251d6e7ec..0000000000
--- a/agent-feedback/unclear.md
+++ /dev/null
@@ -1,21 +0,0 @@
-# Unclear Code & Docs
-
-Things that were hard to understand, and what would have clarified them. Format and rules: [README.md](README.md).
-
-## Index every public docs page in `llms.txt` and add a completeness check
-
-`public/llms.txt` › `Reference` | 2026-07-19 | impact:med | effort:low
-
-`public/llms.txt` is a hand-maintained static index with no generator, and three substantive on-disk pages are linked from none of its topic sections: `docs/reference/lazy-loading.md` (the entire `import ... with { load: "visible#hero" }` client code-splitting / deferred-hydration API, 177 lines), `docs/explanation/class-vs-tags-api.md` ("Tags API for Class API Developers", the flagship Marko 5→6 migration mental-model guide, 245 lines), and `docs/reference/supported-environments.md`. The only aggregate generator, `src/routes/docs/_llms/reference-full%2emd+handler.ts:5-18`, hardcodes a 9-page list that includes lazy-loading but omits supported-environments, so even the bundle llms.txt labels "Complete reference documentation" is itself incomplete, and class-vs-tags-api appears in no aggregate at all. Because llms.txt is an agent's primary topic index — the common pattern is "scan the index, fetch the matching `.md`" rather than pull the large reference-full bundle — an agent asked to add deferred client hydration scans the Reference links, finds no lazy-loading entry, and concludes the capability does not exist even though the doc fully specifies the `load` triggers; an agent porting a Marko 5 class app never discovers class-vs-tags-api.md and mis-migrates from memory. Add the three pages to their sections, add supported-environments to the reference-full list, and add a CI test asserting every `docs/**/*.md` slug appears in llms.txt so new pages cannot silently drop out of the index. Re-verify: `grep -c lazy-loading public/llms.txt` is 0 while `docs/reference/lazy-loading.md` exists (same for the other two).
-
-## Mark `Template.mount`'s required `node` param as required, not `Default: undefined`
-
-`docs/reference/template.md` › `Template.mount(input, node, position?)` | 2026-07-19 | impact:low | effort:low
-
-The `Template.mount(input, node, position?)` parameter table lists a Default of `{}` for `input` and `"beforeend"` for `position` (both real defaults) but `undefined` for `node`, which reads as "node is optional and defaults to undefined". `node` is actually required: `packages/runtime-tags/index.d.ts` in the marko repo declares `mount(input, reference: Node, position?)` with `reference` non-optional, and `packages/runtime-tags/src/dom/template.ts` › `mount` defaults only `input = {}` while `reference` has no default and is dereferenced unconditionally (`parentNode = reference as ParentNode`), so `template.mount(input)` is a type error and throws at runtime. The doc's own examples always pass a node (`template.mount({}, document.body)`), confirming node is mandatory. Agents treat parameter tables as the machine-readable contract: an agent building a client mount or test harness reads "node: Default undefined", writes `template.mount(input)` expecting a detached mount, and burns a debug cycle on the type mismatch — or, working from the rendered doc without a type-check, ships code that crashes on the undefined reference. Replace the `undefined` cell with "required" (or split required params from defaulted ones) so the table agrees with the `reference: Node` signature. Re-verify: read the parameter table under the `Template.mount` heading.
-
-## Document when a `<script>` that reads a tag variable first runs
-
-`docs/reference/core-tag.md` › `## <script>` | 2026-07-30 | impact:med | effort:low
-
-The `<script>` section says the body "is executed first when the template has finished rendering and is mounted in the browser" (line 370) and re-runs when a referenced tag variable changes, which leaves source order as the only ordering a reader can infer between two `<script>` tags. Marko compiles a `<script>` that reads a tag variable into that variable's own setup, so it first runs when the variable initializes, ahead of an earlier-in-source `<script>` that reads nothing: given `<let/board=DEFAULT/>`, a first script that only assigns `board` from `localStorage`, and a second that reads `board` and writes it back, the second runs first with `DEFAULT` and overwrites the saved value before the restore reads it (compiled dom output puts the persist effect inside `_let(...)` while the restore stays a separate `$setup__script`). The page's own `<script=remember/>` example already persists `input.collapsed` to `sessionStorage` without saying when it first fires, so the rule belongs there, stated positively: reading a tag variable ties the effect to that variable's initialization. The paired restore-and-persist idiom belongs under `## <lifecycle>`, whose `onMount`/`onUpdate` handlers already model "set up once, then follow the value" and give `<script>` somewhere to link. Re-verify: `grep -n "finished rendering" docs/reference/core-tag.md` is the page's only ordering statement and `grep -rn "localStorage" docs/` returns nothing, so neither the rule nor the idiom is documented today; the clobber itself reproduces in the local playground (`pnpm run dev`).
diff --git a/cspell.json b/cspell.json
index 1a7ca8d374..d46b43da53 100644
--- a/cspell.json
+++ b/cspell.json
@@ -81,6 +81,7 @@
     "tsrx",
     "TTFB",
     "unitless",
+    "unpursued",
     "untar",
     "ustar",
     "vite",