This PR is quite big, sorry for putting everything into single PR.

Summary

Introduces a canonical, model-agnostic chat IR (crates/spark-server/src/ir/) and migrates every API surface to it, so the core pipeline (build_msg_entries, template rendering, scheduler) consumes one provider-free representation instead of the OpenAI wire structs. Each protocol becomes a thin, unit-tested adapter pair: openai/to_ir.rs / anthropic/to_ir.rs on the request side, openai/encode.rs + encode_stream.rs, anthropic/translate.rs (ir_to_anthropic_response) + the translator.rs streaming state machine, and typed responses_translate.rs / responses_stream.rs on the response side.

Because Message.content is now always a list of parts for every role, the motivating bug — images attached to tool-result messages being silently dropped — falls out for free, along with the rest of the OpenAI-shape tax: the Anthropic struct→JSON→struct round-trip is deleted, the Responses API no longer drops images, and the <think>\n\n</think> and [tool error]\n string-smuggling hacks become real fields (reasoning, an error flag on tool results). The migration commit also fixes a scheduler bug surfaced by end-to-end verification of the tool+image agent loop: the post-think EOS guard fired on every tools-armed turn (eating the model's natural <|im_end|> and causing scaffold garbage plus spurious duplicate tool calls); it is now gated on a sticky think_force_closed flag so it only applies after a watchdog/budget force-injected </think>.

Closes #165 and #164

Test plan

• cargo fmt --all -- --check
• ATLAS_SKIP_BUILD=1 cargo clippy --workspace --tests -- -Dwarnings
• bash scripts/check-license-headers.sh
• Tested against a real model: live on Qwen3.6-35B-A3B-NVFP4 (grammar ON) — /v1/messages turn-2 tool replays 4/4 clean end_turn; full 2-turn tool+image agent loop 3/3 clean (previously 4 turns with chat-template scaffold garbage and duplicate tool calls); new scripts/test-qwen36-tool-image.sh drives the loop end-to-end
• Added or updated tests: cargo test -p spark-server → 689 passed, 0 failed (30 lib + 659 bin). New suites: ir/tests.rs, anthropic/tests/ir_carry.rs (wire→IR→wire carry), anthropic/tests/translator_stream.rs (streaming-framing goldens), expanded openai/tests.rs; prompt-byte stability gates (prompt_json_stability_gate, rendered_prompt_json_matches_retired_json_hop_output) pin rendered prompts to the pre-migration output

Notes for reviewers

• Invariants the tests pin: prompt bytes are the KV-cache prefix and wire framing is what agent clients execute, so two golden gates guard the migration — the rendered-prompt JSON gates (template + Anthropic retired-JSON-hop) and the Anthropic streaming goldens. Any future change to message lowering or stream framing should keep these green.
• Echo-only wire fields (service_tier, metadata, store, include_usage) deliberately do not enter the IR; they ride api/chat/echo.rs::ResponseEcho handler-side.
• Responses API lowers to the chat wire inside openai/ before hitting the IR — intra-provider normalization, deliberate; it now carries images instead of dropping them (from_responses_input_item).
• Accepted wire references outside the adapter edges: response_store (caches wire bodies) and chat_stream --dump synthesis.
• Deferred follow-ups: a neutral ApiError type; top-level reasoning_effort remains dead wire-compat (never wired into the thinking ladder).
• count_tokens now shares api/chat/prepare.rs::prepare_chat_prompt with the serving path (measured end-to-end drift 0.0%).

CLA

• I have read and agree to the Contributor License Agreement.