[Feature] new-work has no rule for a user-requested direction count the verdicts cannot fill, so each run invents one

What problem does this solve?

The hand's size in new-work is an output of the weighing, not a free parameter. skill/reference/new-work.md is precise about the composition: the assigned direction, at most one IMPECCABLE'S PICK card, at most three full-card challengers, and the standing exit. Declined challengers

render demoted, compact and quiet, each carrying its verdict and what the direction kept from it, never full-size, never silently dropped

and dropping a challenger from the hand "takes a named product-truth failure, disclosed."

What happens when the user names a number is not stated anywhere.

My request named five directions. Four of six dealt challengers were declined on both axes — correctly, against a pinned brief of retro fashion editorial, quiet and rich. That left assigned + pick + two competitive = four full cards. The only route to five was to spend a slot on the canon card, which was actively wrong here: the standing exit had already been shown and rejected in round one, and the user's own word for what they wanted was beautiful.

I resolved it by rendering the canon as the fifth screen at full fidelity and disclosing in one line that it is the permanent door rather than a recommendation. That is defensible, but it was my invention. The skill gave me no rule, and the silent failure mode it invites is the opposite choice: quietly upgrading a declined challenger to fill the slot. That is exactly the ranking-rut laundering the verdict system exists to prevent, and it leaves no trace — a promoted card is byte-identical to an honestly competitive one.

Proposed solution

State in new-work that the hand's size follows the verdicts and never a requested number, and give the two permitted moves when a user names a count the weighing cannot fill:

1. Deliver the honest hand and say in one line why it is smaller.
2. Offer a re-roll to deal more candidates that could earn full cards.

Name the forbidden move explicitly — a declined challenger is never promoted to reach a count — because that is the one an agent under a stated number will reach for. And add that the standing exit does not count toward a requested number once the user has seen and rejected it; it stays present as the quiet door.

Worth stating as a general property of the method, not just a patch: when a user names a quantity and the method's own quality gate yields fewer, the gate wins and the shortfall is disclosed. A count is a request about the output, never a licence to lower the bar that produced it.

Provider(s) this applies to

• All providers — the rule lives in skill/reference/new-work.md, upstream of the per-provider transform.

Verified against upstream HEAD cb56ed6 on a clean checkout: the hand-composition and demotion rules are present verbatim; no rule for a requested count exists.

Alternatives considered

• Leave it unstated. This is the status quo, and the cost is that every agent invents a rule under pressure from a stated number. The two inventions are not equally good, and the bad one is undetectable after the fact.
• Cap the request silently. Delivering four with no explanation reads as the skill failing to follow an instruction, which invites the user to re-ask with more force.
• Always re-roll to reach the count. Wrong when the shortfall is the brief doing its job — four of six declined against a tight pin is a signal, not a deal failure.

Related: #817, filed alongside this one. Both are cases where the skill specifies the normal path in careful detail and is silent about the case that does not fit it, so the agent fills the gap. Different surfaces, filed separately.

Willing to work on this?

Yes, if useful — this reads like two or three sentences in the hand-composition paragraph. Per the template I'll wait for your go-ahead before opening a PR.