fix(agent): stop sending the reasoning prefill to native-tools chat providers

[plombeer31]

What

Stops the agent from injecting the local-model reasoning prefill — the trailing <think> open tag (qwen-think) and the Gemma turn-framing tokens — into prompts served over the native-tools chat transport, and makes response parsing stop assuming a prefill that was never sent.

Why

The prefill is a llama-server text-completion artifact: the local template expects the open tag pre-typed at the generation point. On toolTransport: "native_tools" the same prompt ships as a chat message to an OpenAI-compatible endpoint, where the literal tag is at best noise the model echoes back — and at worst corrupted server-side: Ollama Cloud mangles literal <think>/</think> strings in message content (ollama/ollama#17248, ollama/ollama#17617), which is what #283 reports.

Scope note: contrary to the issue's "every request" claim, a pure Ollama-Cloud setup was never affected — think profiles come only from the local llama-server /props probe, so cloud-only runs use plain-instruct and inject nothing. The injection fired in hybrid configs (managed local llama-server probing a think-tag model while completions route to a cloud provider). In those configs there was a second, latent bug this PR also fixes: normalizeContent re-prepended the open tag and the stream parser started pre-opened, so a clean cloud reply that never emitted </think> was silently swallowed whole as reasoning — affecting every cloud provider, not just Ollama.

How

• build-prompt: new suppressReasoningPrefill input drops the trailing reasoning prefill, the Gemma turn framing, and the reasoning system token; the step executor sets it when toolTransport === "native_tools".
• step-executor: promptCarriesReasoningPrefill / completionAssumesOpenReasoning gate every consumer that assumed "the open tag was already sent" — normalizeContent, resolveReasoning, replyFallbackBatch, the stream parser's preOpenedThink, and the one-shot repair prompt (which no longer re-appends <think>). Grammar-parsed completions keep the legacy prepend — the GBNF prelude root structurally starts mid-think — including on cross-transport fallover to a grammar local link.
• profile-invariants: checkProfilePromptAligned learns the suppressed shape (such a prompt must NOT end with a reasoning prelude).
• provider-presets: documents the upstream Ollama Cloud corruption next to the ollama-cloud preset.

What this deliberately does NOT do

The issue's proposed blanket string-replacement of <think>/</think>/【思考】 in all outgoing message content for ollama.com endpoints is not implemented. Silently and lossily rewriting user text and tool results is the same silent-corruption class the issue complains about, relocated into the client — it would break any task where the agent must reproduce file content verbatim. The server-side half is an upstream Ollama serving bug for Ollama to fix; if maintainers still want a client-side mitigation it should be an opt-in provider flag with placeholder round-tripping, i.e. a separate design discussion.

Test evidence

• 5 new vitest cases, all of which fail without the fix (verified by stashing the src changes): qwen/gemma prompt-tail suppression in build-prompt.test.ts; native-tools prompt hygiene + "reply not mis-parsed as reasoning", stream deltas not reclassified as pre-opened reasoning, and repair-prompt hygiene in step-executor.test.ts. Plus a grammar-transport regression case pinning legacy behavior (prefill still sent, reasoning still extracted).
• npm run lint (tsc --noEmit) clean.
• Full suite: 6436 passed / 1 failed — the one failure (src/sidecar/send-message-concurrency.test.ts) fails identically on pristine origin/main (verified in a separate worktree) and is unrelated.

Fixes #283