D5: serve over device KV (4096) returns 500 — refuse with 400 'context exceeds N'; --context-length sizes KV for gate-named cells #4616

OPERATOR ruling D5 (ruling-3715, 2026-09-28), verbatim: "Serve > 4096 tokens → rc500: a 500 is a crash-class lie. Minimum for rc.2: over-length requests return a clear 4xx ("context exceeds 4096"), and serve cells > 4096 are withdrawn by data with a warning, UNLESS gate-named (then fixed: KV size = the model's context)."

Child of #3715: its serve × consumer-context cells above 4096 tokens fail with 500.

Defect: dense CUDA apr serve sends a turn longer than the device KV (max_kv_len(), 4096) to generation. Generation fails with UnsupportedOperation, and the server returns HTTP 500.

Fix (PR #4614):

• serving_context = the model's context, capped by the device KV.
• The chat, completions and stream handlers refuse an over-length turn before generating, with 400 "context exceeds N".
• ContextLimitExceeded maps to 400.
• --context-length N sizes the KV for gate-named cells.

Acceptance:

• A 5432-token turn on lambda returns 500 on main, 400 on the branch, and 200 on the branch with --context-length 32768.
• The unit tests in api/tests/serve_context_d5.rs pass.
• The ladder change for gate-named rungs (pass --context-length) is 0d's fix(#3715 B1): release-readiness ENFORCES at T-1/T-4; report arm deleted (R10) #4603.