Sharpen the keyword index and stop the answer check inventing citations

[devbadya]

A review of what shipped in #18, written as failing tests first. Nine of them failed, which is the whole content of this PR.

Routing: five everyday sentences reached a skill

Message	Routed to	Because
I work out every morning before breakfast	arithmetic	keyword work out
In summary, the trip was a success	summarize-url	keyword summary
Heute war ein wirklich schöner Tag	current-date	keyword heute
Ich rechne damit, dass es morgen klappt	arithmetic	keyword rechne
My sister will find out sooner or later	research-question	trigger find out

Four were keywords an ordinary sentence contains by accident. The last two came from triggers, not keywords, so work out now requires a number somewhere in the message and find out no longer matches after an auxiliary (will, might, to). All five are in the routes-to-nothing corpus.

Dead keywords, and why they were not merely dead. Four skills carried keywords their own triggers already matched — unreachable, and still counted in the document frequency, so they diluted the inverse document frequency of the terms that could be reached. A test now walks every shipped keyword past its skill's own triggers.

carry: false. summarize-url needed a way to decline carry-over: its job is a URL, and carried onto and this one? it handed the model a page reader and no page.

The retrieval index is built once per catalogue rather than rebuilt on every message.

The answer check: three false positives

• A URL is only a citation when something was fetched. Asked what is Anthropic's website, the URL is the answer, and the check was flagging it as an invented source. Both citation checks now stand down when no tool returned a source.
• An exemplar's URL was counted as evidence. The turns the model is sent begin with a skill's worked examples, so a model repeating https://1password.com out of an example instead of citing what it fetched was excused by the very check meant to catch it. Evidence now comes from the real conversation, passed in as options.evidence.
• Exponent values and the typographic minus. 1e+21 has no plain-decimal rendering — toFixed keeps the e — so an answer writing the number out in full read as the wrong number; that check now stands down. And −5 (U+2212, which a model writing prose uses and Number never does) was flagged for disagreeing with -5.

Testing

pnpm check and pnpm build are green: 380 tests, 18 new. Every fix above started as a failing test.

messages the review caught routing somewhere they had no business going
message                                     was                   now
I work out every morning before breakfast   matched "work out"    — none —
In summary, the trip was a success          matched "summary"     — none —
Heute war ein wirklich schöner Tag          matched "heute"       — none —
Ich rechne damit, dass es morgen klappt     matched "rechne"      — none —
My sister will find out sooner or later     matched "find out"    — none —

requests that still have to reach their skill
Work out the tip on 45 euros                arithmetic (trigger)
Berechne 18 Prozent von 2450                arithmetic (search: berechne)
Fasse mir die Seite zusammen                summarize-url (search: seite zusammen)
Find out who won the league                 research-question (trigger)
Wie warm wird es morgen in Rom?             weather (search: wie warm)

carry-over, and the skill that refuses it
and in Lisbon?  (weather resident)          weather (carried-over)
and this one?   (summarize-url resident)    — none —

No UI change, so no new screenshots. Token generation is still not testable here: no WebGPU adapter on the VM and the model has no CPU fallback.