[idpfaff]

The default lazy dispatcher hid tool argument fields and completed 1/20 approval cases against 17/20 for eager schemas. This change makes every profile advertise named, typed tools, removes --tool-discovery with migration guidance, and retains the facade only behind a warned --experimental-lazy switch. Roles and allowlists stay enforced.

The ten-tool API function list drops from 1,485 to 998 bundled o200k_base tokens; instructions grow from 155 to 239, so their combined subtotal drops from 1,640 to 1,237 (24.6%). These are offline serialization counts, not model or billed savings. A snapshot enforces the under-1,000 definition budget. Unknown/missing/type errors now return compact JSON with accepted fields, and SDK-decoded experimental arguments are authorized before execution.

The planned evaluator compares the hash-pinned v0.9.2 eager interface, compact eager schemas, and experimental lazy with typed errors on the same 20 cases. All arms use current execution and identical full JSON/text-only replies; exact numeric grading and historical report publication are covered offline. The command defaults to plan-only and requires --execute for paid calls.

Validation: 1,023 offline tests passed on macOS/Python 3.13; Ruff lint and formatting passed; wheel and sdist built offline. Process-watchdog tests required execution outside the local sandbox. The eager recurring workflow passed 25/25 checks. All seven remote commit trees match their locally tested counterparts. Nothing under benchmarks/results/ or the release workflow was changed.

The corrective v0.10.0 release proceeds through the existing exact-commit full-CI release gate. The eight-job CI matrix passed for the code changes; the final follow-up commit only reconciles release documentation. No paid model evaluation was run. Compact-eager completion and the planned control-minus-one model acceptance check remain unmeasured; this release is an explicit exception to that check, and subsequent default-profile changes must satisfy it.