CL-6484: @corbits/mocks skeleton and Ollama provider

[TheGreatAxios]

Summary

Unit 1 of CL-6484: package skeleton for @corbits/mocks plus the Ollama provider mock. Do not merge — reviewer sign-off requested per task instructions.

• Per-provider entry point (@corbits/mocks/ollama); root export (.) is intentionally empty, existing only so the workspace's check:packages bare-import probe resolves — no barrel that drags providers in.
• Ollama mock covers the three endpoints actually called (@workbench/hub-client's credential-test.ts): GET /api/tags, POST /api/show, POST /v1/chat/completions (tool calls, streaming and non-streaming).
• Request capture + assertions are the highest-value half: expectModel, expectToolsDeclared, expectMessageRoles, expectHistoryContains — the shape of check that would have caught CL-6448 (tools: [], no history reaching the model).
• Usable in-process (mock.fetch as a Fetch-API handler) and as a server (mock.listen()) for the e2e path.
• src/ollama/cl-6448-demo.test.ts demonstrates the assertions failing against CL-6448's exact broken request shape, and passing once tools/history are sent correctly.

Deferred (see README roadmap)

• OpenAI-compatible and Anthropic provider mocks
• Adversarial output catalogue (malformed tool names, truncated JSON, refusals, tool-name-length edges)
• Converging packages/evals' scattered fakes (github MCP fake, stub inference) onto this package

Not verified

• Full repo-wide bun run scripts/checks/packages.ts (packs/consumes every workspace package) was still running in the background when this PR was opened due to its runtime; scoped typecheck/test for packages/mocks and repo-wide bun run lint are green.
• Not tested against a real Ollama instance — this is a mock of the wire contract inferred from credential-test.ts/model-capability.ts, not a live-traffic capture.

Test plan

• bun test in packages/mocks (17 pass)
• bun run typecheck in packages/mocks
• bun run lint from repo root (0 errors)
• Full check:packages repo-wide run (in progress at PR open time)

[TheGreatAxios]

Fixed the FetchLike signature defect from review:

• `OllamaMock.fetch` now mirrors global `fetch`'s `(input: string | Request, init?) => Promise` shape (building the `Request` internally) instead of taking a single `Request`. This is what makes `fetchImpl: ollama.fetch` actually assignable to `@workbench/hub-client`'s `FetchLike` with no cast — the README's own Quickstart now typechecks.
• Routing logic moved to a private `route(request: Request)` method; `listen()`'s `Bun.serve` handler calls that directly so it isn't affected by the public signature change.
• Added `src/ollama/hub-client-consumer.test.ts`, which wires the mock into the real `@workbench/hub-client` functions (`testProviderCredential`, `fetchOllamaModelCatalog` + `preferCompletionCapable`) instead of calling `.fetch` directly — the CL-6477 scenario (embeddinggemma present, chat model correctly selected). This is the test that would have caught the original signature mismatch: it wouldn't typecheck without the fix.
• Removed the stale `expectStream` comment in `mock.ts`.

Scoped `typecheck`/`test` for `packages/mocks` and repo-wide `bun run lint` are green. Repo-wide `check:packages` (packs/consumes every workspace package, including the new `@workbench/hub-client` workspace devDependency) is running to completion before merge — will confirm here.

Still not merging per instructions.