Add Claude Opus: decide the cadence and the budget it needs

[leggetter]

From Gareth, on the first draft:

How much would adding Opus add? That's probably a lot of people's go-to so having
that in the list would be good even if it's similar to Sonnet.

The reasoning is sound: Opus is a common default, and a scoreboard without it invites
"but how does Opus do?" from every reader.

Opus has to run twice, like every other model: once with Hookdeck's skills and
once without, or it cannot appear in the skills comparison the page is built around
(#17). So the unit of cost is a pair of passes, not one.

The figures

A Sonnet 5 pass over fifteen scenarios is about $22, measured. Opus 5 is $5/$25
per million against Sonnet's $3/$15, so an Opus pass is roughly $37, or $55 if
the Sonnet figure was measured during its introductory pricing.

With and without skills, one cycle: $73 to $110.

Cadence	Opus adds	New monthly total
Weekly	+$318 to $477	$503 to $662
Fortnightly	+$159 to $238	$344 to $423
Monthly	+$73 to $110	$258 to $295
Quarterly	+$24 to $37	$209 to $222

Current spend is $185 a month against a $200 ceiling.

Every cadence breaks the ceiling, including quarterly. An earlier version of this
issue said monthly "fits the budget". That was wrong: $185 plus $73 is $258.

So something has to give

1. Raise the ceiling to about $300 and run Opus monthly. The simplest option, and
   a decision rather than an engineering problem. Weekly is indefensible at $500 plus;
   quarterly is too stale to be worth a column.
2. Move the existing weekly to fortnightly, freeing roughly $90 a month, and fund
   Opus monthly from it. Lands near $250. Costs regression-detection speed on the
   models we already track.
3. Drop scenarios. Seven of fifteen pass every experiment. They are a floor and
   worth keeping for that reason, but they are also the cheapest thing to cut. Decide which models and scenarios earn a place in the weekly run #5.

The page already renders a per-row timestamp, so a model on a slower cadence shows an
older date rather than silently implying it was run this week.

Blocked on #14

Adding the most expensive model in the matrix while single-attempt results are
unreliable buys an extra column of numbers we cannot trust. Loop 1 closed as a
negative result for exactly this reason.

Done when Opus is in the matrix at an agreed cadence with the budget to match, or
we have recorded why not.