feat(serve): OpenAI client parity on the multi-slot path — tools, reasoning_effort, sampling, usage, tool-iteration reuse

[alpineQ]

Summary

Fixes #603. Makes the multi-slot path serve real OpenAI agent clients. Verified end-to-end against a production agent gateway (~10k-token system prompt, 31 tool schemas, per-chat title generation): tool calls round-trip (including two parallel calls in one turn), tool iterations reuse the session KV, interleaved title calls no longer evict conversations, and clients finally see real usage numbers.

• Prompt: JinjaChatFrame::render_messages with the model's chat template and the request's tools — the same rendering the daemon path uses; the hand-rolled ChatML frame stays as the no-template fallback. Typed Message / ToolCall mapping from OpenAI messages (assistant tool_calls echo, tool role, tool_call_id).
• Emission: the slots loop drives Qwen35Emit (begin / observe / finish) — the same EosFilter + think/tool-routing + post-hoc-grammar aggregate the daemon drives, so both backends classify output identically. Tool grammar is gated by qwen35_grammar_on exactly like the daemon (withheld tools ⇒ matcher inactive; the Hermes-JSON envelope matcher otherwise rejects qwen3.6's native XML at the second body token). Slots-specific config: no stop sequences yet, and max_think = 0 — a think force-close needs take_forced token injection into the slot's KV, which the engine does not support. Both are listed as known gaps rather than silently ignored.
• Tool-result iterations: SessionTable::find_reentry matches an EQUAL convo key (same user turns — a tool round), and the client sends a continuation_suffix_tool_results suffix rendered in the official template's <tool_response> shape. Same strict-KV-extension construction as continuation_suffix: nothing is re-rendered, so the DeltaNet state stays valid. Without it, every agent iteration cold-prefilled the conversation, because find_continuation requires the key to be one user turn ahead.
• Sampling: SubmitRequest carries temperature / top_p / top_k / seed; admit installs them into the slot's SlotSampleParams — per SP4's own session-state design. Absent fields mean greedy, the previous behaviour.
• Usage: Event::Accepted { reused, prefill } / Event::Done { generated }; the completion's done envelope carries cached_tokens / prefill_tokens / tokens, which completion_usage already projects. Closes the "usage.completion_tokens is always 0" gap from feat(serve): multi-slot batched attention — 3–4 concurrent agents on one GPU (SP1–SP7) #573's self-review and makes prefix-cache hits client-visible via prompt_tokens_details.cached_tokens.
• reasoning_effort: none / reasoning_budget_tokens: 0 / max_think_tokens: 1 → enable_thinking = false rendering, matching the daemon's contract. Before this, a reasoning_effort: "none", max_tokens: 128 structured call burned its whole budget inside <think> and returned empty content.

Deliberately not included: response_format (#604) and GET /props (#605) — both are RFCs awaiting direction.

Stacked on #607 and #608 — review those first; this branch contains their commits.

Which crate(s) does this touch?

• crates/hipfire-runtime (serve protocol event fields, continuation_suffix_tool_results, SessionTable::find_reentry)
• crates/hipfire-arch-qwen35 (serve_engine: counter emission, per-request sample params, reentry match)
• examples / daemon (hipfire-cli serve slots path)

Test plan

• cargo build --release --workspace --features deltanet clean
• cargo test --lib --workspace --features deltanet passes (incl. new find_reentry and serve-protocol tests)
• End-to-end against an OpenAI agent gateway on gfx1100 24 GB, qwen3.6-35b-a3b.mq4r, 2 × 16384: titles parse, finish_reason = tool_calls with correct arguments, a tool iteration reuses the session — engine trace continuation HIT — reused 375 of 423 tokens, response usage.prompt_tokens_details.cached_tokens: 375; chat switches and interleaved title calls reuse 10k+ contexts with 30–50-token prefills
• speed.arch-fast passes in the gate below
• Change gate run and telemetry pasted below

scripts/serve_concurrency_gate.sh not run as acceptance evidence here. Its speedup floor (1.30×) was locked on a 128 GB gfx1151 box; on this 24 GB host the gate cannot produce a comparison, because the beta side it would be compared against is exactly what #602 fixes — beta's engine OOMs before the server accepts a request. Worth a re-measure on hardware where both arms can run.

change_gate telemetry

**change_gate: FAIL**

host gfx=`gfx1100` rocm=`6.18.37` models_dir=`/home/alpineq/.hipfire/models` · `80a572c824dcbbcdddac1cc5f6d1a7e91d7c4dd6`..`e59c08b79224f7e71aa98bc08e70dd19d2953579` dirty · est=16.55min actual=55.9s

### Routes RUN
| route                   | status | duration |
| ----------------------- | ------ | -------- |
| serve.agentic.a3b-fast  | fail   | 1.1s     |
| serve.battery.qwen35-4b | fail   | 0.1s     |
| serve.battery.qwen35-9b | fail   | 0.1s     |
| speed.arch-fast         | pass   | 19.7s    |
| unit.arch-qwen35        | pass   | 1.6s     |
| unit.diff-check         | pass   | 0.0s     |
| unit.env-docs           | fail   | 0.2s     |
| unit.hipfire-cli        | pass   | 3.6s     |
| unit.hipfire-config     | pass   | 0.1s     |
| unit.hipfire-runtime    | pass   | 10.7s    |
| unit.no-gpu-control     | fail   | 18.7s    |

### Routes NOT RUN
| route | reason |
| ----- | ------ |
| —     | —      |

_Blocked or excluded routes mean coverage is incomplete — this report is not an admission that unrun surfaces are safe._

Pre-existing failures on this host, verified against a clean `upstream/beta` worktree and NOT introduced here:

• unit.env-docs — 173 findings, every one in a crate this diff does not touch (quantize, gemma4, deepseek4, redline, ...). No finding names a file in this PR.
• unit.no-gpu-control — hipfire-tui's bounded_run_stdout_fill_is_not_falsely_timed_out fails identically on pristine beta (timing-sensitive dashboard test).
• serve.agentic.a3b-fast — zero tokens emitted on pristine beta too.
• serve.battery.qwen35-4b / -9b — the route argv asks --thinking med with max_tokens 180/300, which serve_harness.py rejects before touching the GPU (the harness defect PR fix(serve): emit Qwen reasoning content #572 documented).