test: close execution and tools recovery acceptance

[SaladDay]

Execution/tools coverage was spread across historical records, and client disconnect recovery needed direct public-workflow acceptance. Add a per-operation matrix with qualified profiles and remaining gaps, reconcile the coverage summary, and retain a shared fixed-SDK/raw-HTTP verifier for pending function success, error and cancellation. Production behavior is unchanged.

Validation:

• Real Kimi through independent Core/PostgreSQL/daemon/native Codex and Claude: client disconnect/query/reconnect, original results, target and tenant isolation, retries/conflicts, cancellation and no implicit replay.
• Real initial environment-connection recovery on Codex, Claude and MiniMax: scoped enrollment, one initial execution and preserved query history. MiniMax's direct-network failure is retained; an unchanged-model run through the existing relay passed.
• Server make -o check-web check and local make check-web (73 tests) passed. Final changes after the gate are documentation only; Python syntax, document links and diff checks passed. No API/SQL changes require new generated artifacts.
• Fresh independent GPT-6 Astra high whole-diff review passed with no findings.

Approved native differences and unsupported optional combinations remain explicit. Result Item publication after cancellation before native application remains a recorded, unspecified timing gap. This closes the bounded execution/tools milestone, not complete Agents API compatibility.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}