Build your own notes workspace (a personal Notion)

Build the project skeleton for a local notes web app. Requirements:
- Node 20+, Express 4, better-sqlite3, HTML rendered server-side with EJS or template literals, no frontend framework, no build step.
- A db.js module that opens notes.db and creates a pages table: id INTEGER PRIMARY KEY, title TEXT UNIQUE, body TEXT, updated_at TEXT.
- An Express server on port 3000 bound to 127.0.0.1 only, routes: GET / (page list), GET /pages/new, POST /pages, GET /pages/:id.
- A shared layout whose header holds a search form posting to /search (stub the route) plus Home and New page links.
- Config in .env loaded with dotenv; add .env and notes.db to .gitignore.
- Acceptance: npm start serves the page list at http://127.0.0.1:3000 and creating a page persists it across restart.
- Out of scope: auth, hosting config, CSS frameworks.
- Pain warning: better-sqlite3 installs a native binary, if node-gyp errors check your Node version before debugging anything else.

Add a Markdown editor with live preview to the pages app from step 1. Requirements:
- On the edit screen, show the existing textarea on the left and a preview pane on the right inside the current layout.
- Render Markdown in the browser with a vendored copy of marked.min.js served locally, no CDN and no npm browser bundler.
- Debounce input events so the preview rerenders at most every 150 ms while typing.
- Saving posts raw Markdown into pages.body from step 1; never persist rendered HTML.
- Server-side display converts body with marked and strips script and iframe tags before inserting into the page.
- Acceptance: a page with headings, a list, bold text and a code block previews identically after save and reload.
- Out of scope: block dragging, slash commands, formatting toolbars.
- Pain warning: pasting rich text can inject raw HTML into the textarea, prove the sanitizer strips it.

Add [[wiki-style]] links and automatic backlinks to the Markdown pages from steps 1-2. Requirements:
- When rendering a saved page, replace every [[Page Name]] token with a link to /pages/<slug>, using one shared slug function that lowercases and hyphenates titles.
- A wiki link to a missing page opens the edit form with that title prefilled; saving creates it, so links never 404.
- Append a Backlinks section under every rendered page listing pages whose body contains [[Current Title]], queried with LIKE against pages.body.
- Style links whose target does not exist differently so gaps are visible at a glance.
- Acceptance: seed Home, Projects, Ideas with Home linking to the others; Projects' backlinks section shows Home.
- Out of scope: autocomplete dropdowns, graph views.
- Pain warning: titles containing regex characters break naive matching, escape them in the LIKE pattern.

Add simple inline databases to the workspace built in steps 1-3. Requirements:
- A page may contain a fenced block labeled db whose first line names typed columns, types limited to text, number, date, select.
- Store each database as a single JSON document in a new dbs table: id INTEGER PRIMARY KEY, page_id INTEGER REFERENCES pages(id), data TEXT holding {"columns":[{"name,type,options"}],"rows":[]}.
- Page view renders the block as an HTML table with editable cells; cell edits POST to /api/dbs/:id/rows which updates the JSON and returns it.
- Select columns render as dropdowns populated from the column definition.
- Acceptance: a Books page with title(text), rating(number), finished(date), genre(select) and two rows survives a server restart intact.
- Out of scope: filters, formulas, relations between databases.
- Pain warning: two tabs editing one row overwrite each other, last write wins, note that in the README.

Wire full-text search across all pages using SQLite FTS5. Requirements:
- Create virtual table pages_fts with fts5(title, body, content='pages', content_rowid='id') plus AFTER INSERT, UPDATE, DELETE triggers on pages to keep it synced.
- GET /search?q= queries pages_fts with MATCH, falls back to a LIKE scan over pages.body when q is only symbols, renders results with snippet() excerpts and the query terms wrapped in mark tags.
- Point the header search form from step 1 at this route.
- Acceptance: a word appearing in three seeded pages returns all three with excerpts; deleting a page removes it from results.
- Out of scope: fuzzy matching, typo tolerance, indexing database rows.
- Pain warning: the FTS5 tokenizer treats brackets as separators, so [[Page Name]] matches oddly, test quoted phrases against bare words.

Add a nightly export of every page to plain Markdown files. Requirements:
- scripts/export.js reads all rows from pages (step 1) and writes one .md file per page into ./export/, named with the slug function from step 3, with YAML front matter carrying title and updated_at.
- Rewrite [[Wiki Links]] in exported bodies to relative Markdown links targeting the sibling .md files.
- Clear ./export/ at the start of each run, then log the count of files written and the destination path.
- Add npm run export to package.json and README instructions to schedule it nightly via launchd on macOS or cron on Linux.
- Acceptance: running npm run export twice yields byte-identical files, one per page, no leftovers from deleted pages.
- Out of scope: uploading exports anywhere, converting database blocks to CSV.
- Pain warning: titles differing only in case collide on macOS filesystems, append a numeric suffix to duplicates.