fix(orchestrate): restore execution delegation and independent acceptance

[Tryanks]

The provider-neutral workflow described Orchestrate as a collaboration channel without clearly assigning execution ownership. Fable could treat “no useful peer discussion” as a reason to implement the entire task itself. Put the lead's operational responsibilities first: decisions, execution dispatch, and independent acceptance of the actual integrated result.

Small implementation tasks use a concise brief and one executor; peer consultation remains optional and pure Q&A stays with the lead. Preserve cross-provider selection, conditional deferred-tool discovery, user work, and recovery before fallback. Acceptance is a responsibility, not a prescribed checklist: the lead chooses its method, depth, and evidence, and a child report does not constitute acceptance. The workflow asset remains the policy owner; the maintained design contract is updated alongside it.

Validation:

• Passed local cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo build --workspace --locked, and cargo test --workspace --locked on the final asset. Four existing environment/live-service tests remain explicitly ignored.
• Direct cargo-machete 0.9.2 found no unused dependencies. Local cargo machete failed because the injected subcommand token was interpreted as a directory; the CI dependency job must still pass. Mobile/Web and other desktop platforms are covered by CI.
• Six real Fable 5.1 / Claude CLI 2.1.263 prompt probes: two implementation cases and one Q&A case, each with HEAD and the candidate. HEAD dispatched neither implementation; the candidate dispatched both. Both answered Q&A without dispatch. In the misleading-report case, the candidate inspected the actual artifact, detected the missing empty-input behavior, and rejected it. I reviewed the traces, artifacts, and validation logs directly.

Probe limits: one sample per condition, scripted MCP executors and reproduced configuration rather than product E2E; no real worker usage or routing guarantee. The candidate also over-specified numeric formatting, and the unresponsive scripted executor caused extra rounds and an eventual direct correction. The numeric-output oracle was relaxed before candidate runs to accept equivalent integer/float rendering; the missing-behavior check was unchanged. These observations support the intended direction, not a statistical behavior guarantee. No string-presence tests were added as a substitute for model behavior evidence.