video: Kling text2video dialect

[CMGS]

What

• Provider kling speaks Kling's real wire: POST /v1/videos/text2video (model_name, string duration, aspect_ratio, resolution: 1080p riding mode: pro), poll GET /v1/videos/text2video/{id}; the {code, data} envelope's non-zero code on an HTTP 200 is rejected as 502 (same treatment as MiniMax's base_resp); the job is keyed by data.task_id — Kling's top-level request_id is a trace id, the same trap DashScope had. succeed bills the clip's reported seconds from task_result.videos[0].duration.
• The previously-invented sync shape stays as the generic fallback; the default config demonstrates it as vidu-video so both paths keep e2e coverage.
• Kling image-to-video is a different endpoint pair and is not wired — an image param on a kling model answers 400 loudly.

Verification

• cargo fmt --check / clippy -D warnings / cargo test --workspace: 541 passed / 0 failed. New: Kling submit-shape assertions (model_name, string duration, no model key, text2video URL), Kling tuple in the dialect e2e (task-id keying, units billed once, 409 before done).
• Live (a 100-point trial package, burned to exhaustion): 13 generations settled — kling-v1-6 std/pro(1080p) at 5 s and 10 s, kling-v2-master at 5 s and 10 s — every clip billed once at the vendor-reported duration (5 s requests render 5.1 s → 6 units; 10 s → 10.6 s → 11 units), 128 billed seconds total. Two 5-way concurrent waves settled every job exactly once — no double billing under concurrency (ledger audited per task id). A transient Kling internal error (task_status: failed, 0 deduction) billed nothing; the exhausted package surfaces as a clean 429 (Account balance not enough). During the concurrent waves the same gateway served 107k rps /health and 114k rps /v1/models at p99 ≤ 1 ms — the async video brokering adds nothing to the hot path.