refactor(sse): narrow three media-generation result unions

[backryun]

Part of #8484. Seven diagnostics across five files, one cause — and the three unions get three different treatments, which is the interesting part.

The cause, again

strictNullChecks: false means ok: true | false narrows the positive branch but leaves the negative one as the whole union. So if (!result.ok) cannot reach status/error. Fifth time this campaign (#8483, #8499, #8531, #8638).

Three unions, three treatments

union	visibility	treatment	fixed
SegmindRequestResult	exported	type predicate	4 (image + video providers)
downloadGeneratedImage's return	module-private	retag	2 (freepik)
ValidatedMediaGenerationBody	module-private	retag	1 (videos route)

Per the rule recorded on #8499 — predicate when exported, retag when private. SegmindRequestResult is the published return contract of segmindRequest(), so its ok shape stays put and gets:

export function isSegmindFailure(
  result: SegmindRequestResult
): result is Extract<SegmindRequestResult, { ok: false }> {
  return !result.ok;
}

The other two have no consumer outside their module, so retagging is cheaper and clearer. Both consumers of the retagged body union — the videos and music routes — were updated with it.

What I tried, measured, and reverted

MediaGenerationResult in the same shared file looks like the same shape:

type MediaGenerationResult =
  | { success: true; data: unknown }
  | { success: false; error: unknown; status: number };

Retagging it and narrowing failedMediaGenerationResponse's parameter to the failed arm fixes 2 more — but introduces one new error: the provider handlers that produce these results widen success to boolean rather than a literal, so the route's result no longer matches the narrowed parameter.

That is a different root cause — literal widening at the producers, not discriminant narrowing at the consumer — and fixing it properly means going after the producers first. Measured 10 fixed / 1 new, reverted that part, kept the invariant. Left for its own slice.

Verification

208 → 201, zero new errors, on a line-number-agnostic diff of the full tsc error set.

• npm run typecheck:core — clean
• eslint on all seven changed files — clean
• check:file-size — clean
• 234/234 across the 27 affected suites

No new tests — the failure arms are exercised

Checking the arms the change touches is standard here, and both retagged/predicated failure paths already run under test:

• segmind-image-video-provider-6656.test.ts drives a 500 (text/plain body) and a 502 through both segmind providers — the exact ok: false path isSegmindFailure now guards
• video-generation-handler.test.ts covers a 503 and a malformed-output 500

The freepik download failure and the invalid-JSON body arm are narrower; they are reachable and unchanged in behaviour (retagging a discriminant has no runtime effect beyond the literal string compared), so I did not manufacture coverage for them.

[diegosouzapw]

Validated in local merge-train /tmp/train1d-20260727-090022-suite.log on .113 @ 029cdf4215cf465f0e1716ac9f84a84692b1e881 (full unit suite green on CI-equivalent host)