[inureyes]

Summary

• Preserve template-rejection sentinel errors through chat request preparation so downstream routes can distinguish deliberate template refusals from engine-side render failures.
• Return HTTP 400 invalid_request_error from the disaggregated router when a chat template rejects caller-supplied values.
• Make offline CLI prompt resolution return a named error for template rejections while retaining bare-prompt fallback for non-rejection render failures.
• Add committed English and Korean technical reports for this open PR because TECHNICAL_REPORTS/.keep-reports is present.

Validation

• PASS: cargo test --bin mlxcel resolve_cli_prompt
• PASS: cargo test --bin mlxcel apply_user_chat_template_wraps_prompt_as_user_message
• PASS: cargo test --bin mlxcel vlm_chat_template
• PASS: cargo test --lib responses_reasoning_effort_template_rejection_survives_translation
• PASS: cargo test --lib responses_template_engine_failure_still_uses_prompt_fallback
• PASS: cargo test --lib router_chat_maps_template_rejection_to_bad_request
• PASS: cargo test --lib template_rejection
• PASS: cargo clippy --lib --tests -- -D warnings
• PASS: cargo clippy --bin mlxcel -- -D warnings
• PASS: hosted checks on report commit 998a70b: cargo-clippy, cargo-deny, cargo-fmt, OpenXLA feature compile, crate versions, cross-repo refs, kernel dtype keys, llama-compat manifest, Detect changes, and license/cla. MLX pin extraction and OpenXLA feature link skipped by change detection.
• SKIP: broad cargo workspace tests and cold release builds per watchdog guard.

Reports

• TECHNICAL_REPORTS/1511-template-rejection-surfaces-20260830.en.md
• TECHNICAL_REPORTS/1511-template-rejection-surfaces-20260830.ko.md

Closes #1176