spaces: page aliases and unlinked mentions

[nyblnet]

Two features that are one idea: a page is known by more than one name, and those names appear in prose whether or not anybody typed [[ ]] around them. New logic is confined to spaces/src/mentions.ts; the diffs against the files nine sibling branches are also touching are small (editor.ts +110, model.ts +23, main.ts +13, agent.ts +42, props.ts +40, markdown.ts untouched).

Aliases

Page.aliases — ["NYC", "the Big Apple"] on a page titled New York — edited in the page panel as one comma-separated field.

Resolved at link time, in one function (nameIndex), read by the [[…]] resolver, ⌘K and the [[ page picker. What gets written into the file is an ordinary #p/<id> href, so backlinks, the graph, export, print and the CRDT never learn that aliases exist. All three surfaces or none — an alias that links but cannot be searched means you file something under the name you use for it and then cannot find it by that name, which reads as the search being broken.

Additive: absent by default, and clearing the last alias deletes the key rather than storing [], so a page that had one and lost it is byte-identical to one that never had one. Verified through the shell's own serializer.

Collisions are reported, never repaired. A title always beats an alias, then document order — identical in every replica. validate() emits alias-collision naming the page a [[link]] will actually reach.

Unlinked mentions

Under the backlinks: every place this page's title or aliases appear as plain words elsewhere, with the sentence and a Link button that wraps exactly the words that matched.

Most of the work is what it refuses to find — a title inside a code block, a code span, an existing link, a URL or a mail address; a block that already links to the target; the page's own text; part of a longer word; anything under three characters (two for Han/kana/Hangul).

The matching rule, and CJK

The obvious boundary is (?<![\p{L}\p{N}_]). Applied to Japanese it does not find fewer mentions — it finds none, ever, because every kana beside a name is a letter and the assertion never opens. That is a silent total failure in three of this app's nine locales, and it passes every test written in English.

So a boundary is required only where the name's own edge character is a word character in a script that separates words. 私は東京に住んでいます mentions 東京. The stated cost, because it is real: a Han name also matches inside a longer Han compound (京都 inside 東京都). Both halves are asserted.

Two mechanical notes: the v flag expresses this as one set difference and is not used — v is Safari 17 against this app's Safari 16.4 floor (DecompressionStream), and a regex the engine cannot compile throws on every page open rather than missing a match. And the run scanner separates skipped regions with U+0000, not a space: with a space, New<a>x</a>York collapses to New York and reports a mention nobody wrote.

Cost — measured, not assumed

"Every page's title against every page's text" is N×M and the panel never computes it. Opening a page scans the document once for that page's names, so cost is the size of the space, not the number of pages in it.

synthetic space	buildIndex	one page's mentions	all pairs
100 pages, 252KB	0.2ms	1.3ms	62ms
500 pages, 1.26MB	1.1ms	3.4ms	417ms
1000 pages, 2.5MB	1.9ms	6.5ms	1000ms

The all-pairs column is bento.mentions(), the read-only agent surface, and is on no paint path. One measurement corrected a claim in an earlier draft of these comments: the four ways of spelling the boundary all run within noise of each other (0.03–0.05ms per 210KB pass), so the choice is the browser floor and nothing else. The real cost I had introduced was an unanchored email pattern in the URL mask — 7ms → 21ms — now gated.

Shell: +5.0KB compressed (277,769 → 282,781 B), measured by building origin/main and this branch back to back. The size rig reports +24KB against a reference that has drifted behind other merged work.

Verification

• node scripts/test-spaces.mjs — all 9 rigs pass (new mentions rig: 59 assertions); --manifest ok
• tsc -b clean; build:single + shell-gate clean; test-sync-spaces 3/3
• i18n: 7 new strings, all 8 catalogs, packed.ts grepped — 100% every locale
• 13 sabotages, each guard removed in turn and watched to go red, then restored. The self-mention guard is doubled, so removing either half proves nothing — both had to go before the rig noticed, and that is now written next to the code.
• Browser-verified on a uniquely-named copy over http://127.0.0.1:5417, marker-checked (typeof bento.mentions === 'function'). All five negatives held in the real shell; the Link button spliced exactly the right occurrence and the mention became a backlink; ⌘K and the [[ picker both found the page by alias with real keystrokes — synthetic input events do not reach the [[ handler, which is exactly the instrument that lies. No console errors.

Notes for review

• .github/workflows/ci.yml and scripts/test-spaces.mjs are the shared surface, claimed by path. Not optional: test-spaces.mjs exits 2 on a rig that exists but is unlisted, and --manifest exits 2 on one listed but unregistered. One step, appended at the tail.
• scripts/test-spaces-mentions.ts is new rather than an extension of test-spaces-model.ts, to keep out of a contended file.
• docs/DECISIONS.md currently ends with a Claude-Session: URL from an earlier session. Left alone — not mine to rewrite — but flagging it, since it is the attribution the standing rule forbids.

[github-actions]

Build size

main (0100083) → spaces-mentions (b006413)

app	base	PR	change
bento/slides	677.9 KiB	677.9 KiB	0.0 KiB (0.00%)
bento/spaces	271.3 KiB	276.2 KiB	+4.9 KiB (+1.80%)
bento/dash	424.1 KiB	424.1 KiB	0.0 KiB (0.00%)

Updated: 2026-09-10T00:22:50Z