fix: the fifth leaking branch, and a claim that outran its code again

[fstubner]

A verification pass on #215 returned NOT CONFIRMED.

#215 asserted that "every branch that returns endpoint-authored text now goes through one #safe()". A fifth branch did not:

error: streamError ?? this.#safe(describeUnusableBody(rawSeen))

The scrub sat on the fallback. streamError holds the endpoint's own message, parsed straight out of an SSE error event by both wire-protocol parsers, and it reached DispatchResult.error — and logs/dispatches.jsonl — verbatim.

It is also the likeliest shape of the lot: a 200 with text/event-stream carrying an error object is how OpenAI-compatible gateways and Anthropic both report an auth failure once the stream is open, and jobs only ever stream. Reproduced against the built artifact with a URL-embedded key and one echoed from the Authorization header.

The claim, not just the code

This is the second commit message about this file to describe a wider fix than the code delivered. The #safe docblock no longer asserts completeness — it enumerates the branches that assign DispatchResult.error and says which are ours and which are scrubbed. An enumeration can be checked against the file; a claim of completeness cannot.

Also fixed: describeFetchFailure has always scrubbed the base URL but was never given the api key — the last unscrubbed-for-key site in the file.

Sabotage

Reverted	Fails
#safe moved back onto the fallback	exactly the 2 mid-stream tests
the key argument to describeFetchFailure	exactly the fetch-failure test

Verified at the artifact

All seven error branches driven against built dist/ with a stubbed fetch, no network:

ok   mid-stream sse error (openai)      ok   unusable 200 body (buffered)
ok   mid-stream sse error (anthropic)   ok   http 400 (buffered)
ok   unusable 200 body (stream)         ok   http 400 (stream)
ok   fetch failure

npm run check: 1226 passed, 8 skipped, exit 0.