feat(discovery): list a custom endpoint's own models in the picker - #287

The picker never asked a custom endpoint what it serves

Scoping the model picker to a customEndpoints provider opens an empty list. The
provider row reads 0 models, and no notice says why.

The endpoint works — --model my-endpoint@some-id "task" reaches it and streams. The id
just has to be typed from memory, because nothing ever asked the host for its dynamic
models catalog.

Cause

buildProviderDefinition (providers/custom-endpoints-loader.ts) set no modelDiscovery
descriptor. PickerDataSource.providerList() therefore reported hasDiscovery: false for
every customEndpoints entry and for every one of the 27 bundled predefined-catalog.ts
rows, which compile to CustomEndpointComplex and travel the same function. ModelPicker's
list memo then falls back to catalog-scoped rows — empty for a private host — and
providerCountText prints 0 models.

rows.tsx:315's other phrasing, lists its own models, was written for exactly this case
("a self-hosted vLLM") and was unreachable, since that branch needs hasDiscovery.

What this changes

An endpoint that declares the openai wire format now derives a descriptor: the /models
path beside the /chat/completions it already declares.

Derived, never configured — no new config surface. format: "openai" is already the
user's statement that the host speaks the OpenAI API, and GET /models is part of that API,
published as a sibling of /chat/completions. So this is one rule applied to the path the
entry already carries, not a pinned URL — which is the distinction the "never hardcode a
provider's models" invariant draws. All 27 bundled rows declare only /v1/chat/completions
or /chat/completions (the version segment lives in baseUrl), so each lands on the path
its vendor actually serves.

Only openai. The anthropic transport's /v1/models wants an anthropic-version header
and answers created_at where openai-models-list reads created; litellm and gemini
are different shapes again. Each deserves its own change with its own live verification.

Discovery stays on demand — it runs when the user scopes to that provider, never at
startup. An endpoint that does not serve the path becomes a failed outcome naming the URL
and the status, which is strictly more than the silent empty list it replaces.

Three commits, separable

1. feat(discovery) — the derivation, its tests, and the docs for it.
2. fix(discovery): stop reporting no-credentials for authScheme "none" — a keyless
   endpoint (Claudish configuration for Clawrouter #139) reaches discoverProviderModelsCatalog with static headers and no auth
   header by design (api-key-credential.ts:362). The blank-credential gate could not tell
   that from an unset key, so it would have sent the user hunting for a variable they
   correctly do not have. One line; reachable only because of commit 1.
3. docs(endpoints) — both docs said the models field "restricts the endpoint to
   listed models". No code has ever enforced that: its only consumer is discoverableModelIds
   in proxy-server.ts, which builds the proxy's own /v1/models — the list the nested
   Claude Code's /model command reads. Drop this commit independently if you'd rather keep
   the wording.

Verification

• Live, against an OpenAI-compatible endpoint whose /v1/models this code had never
  requested: hasDiscovery: true, outcome served in 1164 ms, 427 models, 140
  chat-capable after isReportedChatCapable — including ids previously reachable only by
  typing them.
• bun test packages/cli/src/providers packages/cli/src/picker — 1442 pass, 1 skip, 0 fail.
• Mutation-checked: reverting the commit-2 line fails its new test with Received: "failed".
• bun run typecheck clean; bun run format no fixes; bun run lint exits 0.

One note on lint: biome reports cognitive complexity 33 on discoverProviderModelsCatalog.
The baseline there is already 32 — the authScheme check adds one — and it is one of 22
pre-existing warnings repo-wide, so I left it rather than refactor a function this PR only
touches by a line.