[senamakel]

What

Lets an embedder set the wording a turn is closed with when it stops at its tool-call cap or ends without a reply (CloseCopy: the model's cap instruction, the final-write instruction, the deterministic fallbacks and the repair prompts).

• RuntimeBuilder::close_copy(CloseCopy) sets it per runtime; Turn::close_copy(CloseCopy) overrides it for one turn. CloseCopy and ENGLISH_CLOSE_COPY are re-exported from openhuman-embed.
• It is carried on HostTurnEnv / OpenHumanRunContext (inherited by children) and read in one place, OpenHumanRunContext::close_copy(), which both context_wiring (FinalCallWrapUpMiddleware) and grounded_close / execute now take instead of the hard-wired OPENHUMAN_CLOSE_COPY. The two paths cannot drift apart.
• product_copy::close_needles adds the instructions actually in force to the needle list, so the echo-strip still removes a reworded instruction the model repeats; is_final_summary recognises the copy's own "finished without a result" opening instead of a hard-coded English sentence.

The default stays ENGLISH_CLOSE_COPY, so OpenHuman's own behaviour is unchanged.

Cause: the close copy was a process-wide constant, so a host whose turns are chat replies to end users could not stop a capped turn from being told to write an operator report.

Closes #702
Part of tinyhumansai/chottu-core#537

Tests

• an_embedder_close_copy_reaches_the_capped_turn (runtime copy and per-turn override reach the model and the checkpoint reply), the_default_close_copy_is_unchanged (openhuman-embed/tests/close_copy*.rs)
• a_reworded_close_instruction_echoed_by_the_model_is_recognised and related product_copy tests (core)

Gates and caveats

fork/epic/537-turn-quality does not compile openhuman-core on its own (21 errors, from the merged tinyhumansai#8487 state: e.g. secure_open path, await in non-async fn), so the gates were run on a throwaway local branch = this branch + fork/pr-8487-followup (not pushed):

• cargo test -p openhuman-embed --features inference,mcp,skills --test close_copy --test close_copy_default: pass, but only with the prompt screen bypassed locally. The security module artifact is not installed in this sandbox, so the screen holds every prompt; sibling embed turn tests (agent_lifecycle, observed_turns) fail identically here with the screen intact. Needs a CI run with the module fixture.
• cargo test -p openhuman --lib -- agent::session_host::product_copy agent::session_host::driver::grounded_close agent::tinyagents::context: 50 pass, 2 fail (the_final_summary_of_a_halted_turn_gives_a_reason_and_names_the_tools, classified_halt_returns_partial_work_without_provider_usage); both assert on tool error text that arrives as [REDACTED] from the followup state and go through code this PR does not touch.
• cargo clippy -p openhuman -p openhuman-embed --all-targets -- -D warnings: no diagnostics in files changed here (existing unused-import errors elsewhere).
• cargo fmt -- --check: clean.

Review round 1 follow-up

• Added wrap-up coverage (wrap_up_close_copy, max_iterations >= 3): the penultimate request carries the custom final_write_instruction, the last permitted one the custom max_iteration_instruction, no English default.
• close_needles now includes stop_note_preamble; grounded_close test pins the echo repair with a reworded copy.
• The host policy resolves the default once via OpenHumanRunContext::close_copy().
• builder-setters / api-index docs regenerated.
• The per-turn override lives on Turn rather than TurnRequest on purpose: TurnRequest is the serialized inference.agent_chat wire contract and cannot carry copy.
• Verification: the embed tests run only on a throwaway branch (this branch + fork/pr-8487-followup + a local prompt-screen bypass), because openhuman-core does not compile on the epic tip; two core tests (the_final_summary_of_a_halted_turn..., classified_halt_returns_partial_work...) fail identically on the base. The branch carries no bypass and net-zero submodule/Cargo.lock change.