Truncated model replies are never executable

A completion cut off by max-tokens or a broken stream can end mid-<codemode> block or mid-fence; both interpreters (the platform's AgentCodemode and codemode-tag's vendored parser) will happily extract and run whatever half-script survived.

This was fully implemented and spec'd on the prompt-sections branch, then deliberately reverted before merge to keep PR #2512's diff to the prompt model itself. Start by reapplying commit a9e3c7ef3 (git cherry-pick onto main after #2512 merges; expect light conflicts in the two test files and regenerate itx-api). Its commit message is the design record. Summary:

• SSE drain in workers-ai-transport captures choices[0].finish_reason (last non-null wins — it rides the empty-delta final frame); non-streamed path reads it off the response object; journaled in rawResponse evidence
• AgentLlmTransport callLlm seam gains optional finishReason so harness transports can simulate truncation; absent = assume complete (dialects that never report one fail open, not block every reply)
• assistant agents/context-added payload gains optional truncated: boolean (on the EVENT: userland workers see only events, and replays must see the fact)
• both interpreters guard before parsing: corrective developer feedback ("your reply was truncated — NOTHING was executed; re-send"), after-current-request, idempotency key agent/truncated-reply-rejected@<offset> mirrored in both so the birth race dedupes (same pattern as malformed-snippet feedback)
• specs: transport finish_reason ("stop" on the existing gpt-5.5 fixture, "length" with rawResponse evidence), full harness loop (truncated reply → no script-run-requested → feedback → retry executes), and the parsing-off path (marker event-visible, platform interprets nothing)

Resolved question, for the record: AI Gateway BYOK proxies chunk bodies verbatim and the response cache replays recorded SSE byte-identically; finish_reason lives in chunk JSON, not headers — nothing strips it.