Replace the openai SDK with a lean fetch-based client.

[expilu]

Drop the openai dependency — lean fetch-based generateText client

Summary

smart-decisions now talks to OpenAI-compatible endpoints through ~100 lines of its own
transport code instead of the openai SDK. The library ships zero runtime dependencies.

The public API is unchanged: choice() keeps its shape and signature. Question gains two
optional provider settings: maxRetries (default 2) and timeoutMs (default 600000).

What changed

New transport — src/utils/llms/generate-text.ts

• Single non-streaming POST {apiBaseUrl}/chat/completions with Authorization: Bearer and a
  JSON body including stream: false — the same wire format the SDK produced (verified by
  capturing the exact request).
• Configurable retries (default 2, SDK-style counting: total attempts = maxRetries + 1) with
  exponential backoff (500ms doubling, capped at 8s), honoring retry-after-ms / retry-after
  and the x-should-retry header. Retryable statuses: 408, 409, 429, ≥ 500.
• Per-attempt timeout (default 10 min) via AbortSignal.timeout. Timeouts are deliberately
  not retried — doubling a 10-minute wait for a one-token decision is worse than failing fast.
• Errors are plain Errors carrying the HTTP status, a truncated response body, and the
  original failure as cause.
• User-Agent: smart-decisions/<version> (src/version.ts, kept in sync with package.json
  by a test).

Repo structure

• src/utils/{time,network,error,text}/ — eight focused helpers (sleep, backoffDelay,
  numericHeader, retryDelayFromHeaders, shouldRetryStatus, isTimeoutError, messageOf,
  truncate).
• src/types/ — seven granular internal types describing the chat-completions
  request/response shapes.
• House doc standard applied repo-wide: self-documenting JSDoc with @param, @returns,
  @throws and @example on every exported symbol.

Tooling

• tsconfig.json is now a repo-wide typecheck project (src + test + examples, noEmit);
  emitting lives in tsconfig.build.json. New typecheck script, wired into prepublishOnly.
  Tests and expectTypeOf assertions are now enforced by the compiler.
• .vscode watch task now runs the full-repo typecheck (no dist/ churn on save).

Tests

• The transport is tested through the global fetch with real Response objects and fake
  timers; system1-choice tests re-wired the same way; new unit suites for every helper and
  type; a version-sync test.
• 23 files / 69 tests, 100% lines / branches / functions / statements (thresholds enforced).

Behavior deltas vs the SDK

	openai SDK (before)	now
Retries	2 fixed, incl. timeouts	configurable (default 2), timeouts excluded
Timeout	10 min per request	configurable, 10 min per attempt (default)
Errors	APIError subclasses	plain Error (status + body + cause)
Headers	OpenAI/JS x.y.z, X-Stainless-* telemetry	smart-decisions/x.y.z, nothing else
Dependencies	openai	none

Verification

• typecheck, lint, format:check, build, test with enforced 100% coverage
• Live end-to-end: npm run example against llama.cpp (movie @ 0.9991, matches the
  README's documented numbers)
• Old-SDK vs new-transport equivalence on the live server: same answer and same token
  ranking; the small logprob deltas were proven to be server-side run-to-run noise via a
  same-client control (two identical requests through one transport also wobble)
• Captured wire request: identical method, path, headers and body contract
• npm pack — 51 files, 47.3 kB, zero openai references in dist or lockfile