feat: new-since-last-visit filter with keyboard-reachable filters

[ericboehs]

What

Two things, both about getting through the listing faster.

New since last visit

The listing marks stories that were not there on your previous load, and a pill under the score filter narrows the page to just those.

It runs in the browser on purpose: the pages are served public with one ETag shared by every visitor, so per-visitor state cannot live on the server without sessions or giving up caching.

• Baseline is a set of story ids, not a timestamp. A story is often hours old by the time its score lifts it into the listing; filtering on when it was posted would hide exactly those arrivals.
• Five-minute grace window. Refreshes in quick succession keep the same marks rather than clearing them before they have been read.
• Bounded state. The seen map lives in localStorage under hn.visit.v1 and expires after seven days.
• No dead controls. The pill ships hidden and is revealed only after the script confirms it can write to storage, so a visitor without JS or in Safari private mode never sees a toggle that does nothing.

Date dividers are now rebuilt from the visible rows on every filter change. Inserting them once left them stranded above stories from another day as soon as rows were hidden; the rebuild also retires the old textContent.includes dedupe hack.

Keyboard-reachable filters

k past the first story steps up into the filters — the new-since pill, then the score pills, landing on the score already in effect — where h/l walks the row and space applies it. j walks back down.

• h/l is clamped inside a filter row, so stepping past 0 or 1000 cannot drop focus into the page heading behind it.
• Hidden rows are skipped, and the walk no longer throws at the ends of the list.
• ? opens a cheatsheet listing the keys each page actually binds (the listing advertises the filter shortcut, story pages do not).
• UIkit strips the focus ring from subnav links, which left the whole walk invisible; restored for keyboard focus only.

Testing

bundle exec rake test — 127 runs, 0 failures. The Ruby tests are structural guards (control ships hidden, row ids present, nav hooks intact); behaviour was driven end-to-end in headless Chrome over CDP with real key events, in both colour schemes:

• first visit marks nothing, a later load marks only the arrivals, grace window holds across a quick refresh, empty state swaps the table for the placeholder
• j → story, k → toggle, k → active score pill, h h → clamped at 0, space → navigates / toggles, ? → opens, ?/Esc → closes, nav keys stand down while it is open

The README documents a console snippet for exercising the filter without waiting for HN to post something.