Moving to draft — honesty note on the evidence.

This change is a defensible contract-alignment (deepseek-v4-pro genuinely is a thinking-with-tools model: its assistant turns carry thinking + tool_use blocks), and the original exclusion rested on the now-questionable "#4842 works out-of-the-box" assumption.

But I can't currently confirm it fixes a problem that manifests in normal use. The incident that prompted it was an agent on deepseek-v4-pro 400ing with "reasoning_content must be passed back". On closer inspection a sibling agent on the same model runs 100+ thinking + tool_use assistant turns without the echo policy and never hits the 400 — so the missing echo policy can't be the whole story. The real differentiator between the two agents is max_history (aggressive trimming on the failing one), not the echo policy.

Holding as draft until I can confirm whether deepseek-v4-pro actually requires the echo (and isn't getting it some other way for the working agent). Did not want to put an unconfirmed fix in the review queue.

Confirmed live — promoting out of draft.

Deployed this fix to a host running a deepseek-v4-pro agent that was 400ing on every tool-call continuation with "reasoning_content must be passed back". After the deploy the agent runs its multi-turn tool-calling task cleanly — no 400, no agent-loop failure, reasoning_content round-tripped on each turn:

usage prompt_tokens=23238 completion_tokens=100   ✓
Captured reasoning_content from response len=1417  → completion_tokens=696  ✓
completion_tokens=3185                              ✓
Captured reasoning_content len=4852                 ✓

So deepseek-v4-pro does require the same reasoning_content echo as V4 Flash; treating it as thinking-with-tools fixes it. (The sibling agent on the same model that appeared to work without the echo turned out to be a red herring for the fix — both echo it now and both run fine.)

Reviewed — the change is correct, minimal, and well-tested. Two non-blocking notes.

1. Fix targets the fallback, not the catalog — confirm the policy actually reaches the wire.
fallback_reasoning_echo_policy only fires when the request carries no explicit policy (ReasoningEchoPolicy::None — catalog miss or pre-registry build), per its own doc comment.
If a community-catalog entry for deepseek-v4-pro ever sets a non-None reasoning_echo_policy, the request carries that value and this fallback never runs, making the change a no-op for catalog-resolved requests.
The added test verifies is_deepseek_v4_thinking_with_tools("deepseek-v4-pro") == true at the function level, but not end-to-end that Echo lands on the actual outbound request.
Low risk in practice: there is no in-tree catalog entry for the model (builtins sync from the external registry into home_dir/providers/, default None), and the production 400 was observed on exactly this fallback path — so the fix targets the right layer.
Flagging only so a follow-up that adds a catalog entry doesn't silently shadow this.

2. Breadth check — no parallel site needs the same change.
The other model-aware thinking path, parse_labeled_summaries / strip_think_preamble in crates/librefang-runtime/src/history_fold.rs, is model-agnostic (it strips any leading <think>…</think> block; deepseek-v4-flash is only named as an example, not gated on). So is_deepseek_v4_thinking_with_tools is the only model-gated point, and this PR covers it.

Nit: the CHANGELOG bullet carries only (@DaBlitzStein) — sibling entries in the same ### Fixed block include the issue/PR number too (e.g. (#6083)). Consider appending (#6097) for consistency.