engines: carry the parallel tool policy to Claude models on Converse

[CMGS]

What

Follow-up to #94. On Converse a chat client's parallel_tool_calls stayed in additionalModelRequestFields because toolConfig has no parallel flag; the OpenAI family accepts it there, a Claude model answers 400 "parallel_tool_calls: Extra inputs are not permitted" — while the same request succeeds on anthropic-messages and aws-anthropic.

Bedrock accepts the whole tool_choice (with disable_parallel_tool_use) in additionalModelRequestFields as long as toolConfig.toolChoice is absent, and refuses the pair ("The additional field tool_choice/type conflicts with the existing field toolConfig.toolChoice.auto"); a flat disable_parallel_tool_use is refused too. So for a Claude model the policy becomes tool_choice.disable_parallel_tool_use and that choice rides in the extras instead of toolConfig; other families keep parallel_tool_calls there unchanged.

• converse::request: claude_model(model) (the existing contains("claude") check, now shared) gates the routing; one line of comment carries the vendor constraint.
• claude_engine: the parallel_tool_calls → disable_parallel_tool_use mapping runs on Converse too when the model is Claude.
• tests: the Converse integration test now walks both families × both policies; a unit test pins the routing and the OpenAI-family passthrough.
• docs: API page states the Converse behavior per family.

Hot path

Zero increment by construction: the family check reuses the one already computed in request, and a routed choice is one fewer move (it stays in the body that becomes additionalModelRequestFields). Requests without tool_choice, and every non-Converse wire, take the same path as before.

A/B bench on the Linux testbed (384 cores), mock upstream, oha -c 256 -z 15s per shape, six rounds with the arms interleaved and the order swapped each round. Arm A = main before #94 (a66250a), arm B = this head, so the table covers the #94 paths and this one:

shape	A rps (median)	B rps (median)	Δ rps	A p50	B p50	Δ p50	paired Δ rps per round
chat, no tools (untouched path)	237884	240585	+1.1%	0.358 ms	0.360 ms	+0.7%	+2 +5 −1 −4 −1 +6
chat + tools + tool_choice + parallel_tool_calls → openai-chat	217502	208700	−4.0%	0.339 ms	0.329 ms	−2.8%	−22 +14 −6 0 −2 −12
/v1/messages + tools + {type: any, disable_parallel_tool_use} → openai-chat	162964	164300	+0.8%	0.338 ms	0.342 ms	+1.0%	−19 −17 +27 +42 +7 −15
chat + tools + parallel_tool_calls → anthropic-messages	263682	271192	+2.8%	0.348 ms	0.362 ms	+4.0%	+9 −30 +4 +25 +1 +48
chat + tools + parallel_tool_calls → aws-converse (this PR)	159053	152609	−4.1%	0.312 ms	0.307 ms	−1.4%	+11 −6 −10 −23 +20 −14

Every per-round paired delta changes sign across the six rounds and the medians sit within ±4% with rps and p50 disagreeing on direction, so the run resolves nothing below its own ±10–40% round-to-round noise. Success rate 1.000 on every shape and arm.

The two shapes with a −4% median were re-run at -c 64 (noise sd ≈3%), eight rounds, and then once more as an A/A control with the baseline binary on both arms:

shape	A/B mean Δ rps (8 rounds)	A/A control mean Δ rps (8 rounds)
chat + tools + tool_choice + parallel_tool_calls → openai-chat	−1.7% (sd 2.6)	−1.7% (sd 3.3)
chat + tools + parallel_tool_calls → aws-converse	+2.0% (sd 3.7)	+1.8% (sd 8.4)

The identical binary on both arms reproduces the same signed offsets, so the residual is the harness's arm-position bias, not code; the source delta on the openai-chat path is two lookups on a three-key map (under 0.1 µs against a 227 µs p50). No regression.

Gates

cargo fmt --check clean; cargo clippy -p gw-engines --all-targets -- -D warnings clean; engines tests 215 passed locally. Linux (rustc 1.98.0, kernel 6.12) cargo test --workspace at this head: 32 suites, 632 passed, 0 failed, 4 ignored (the Redis- and Postgres-gated suites, no credentials on the testbed).

Live (Linux testbed, real Bedrock us-east-1, Gemini, OpenAI, Anthropic: 22/22 on this head, gateway logs 0 WARN/ERROR)

check	before	after
chat parallel_tool_calls: false + tools → Converse us.anthropic.claude-haiku-4-5	424 carrying Bedrock 400 "parallel_tool_calls: Extra inputs are not permitted"	200, one tool call
same → Converse openai.gpt-oss-20b	200	200
same → aws-anthropic InvokeModel global.anthropic.claude-haiku-4-5	200	200
Converse haiku without the field (control)	200	200
the #94 suite (Gemini signatures, DLP replay index, body cap, tool_choice shapes, stop_reason, parallel policy on direct wires)	18/18	18/18

Lines: production +14, tests +33, comments +2 (two one-line docs), docs +5/−3.