sdk: harden the chain provider - bound block_on and guard ChainMethod drift

[mfw78]

From review of #453 (comment 1, comment 2). Confirmed still valid at HEAD. Target: L1 nexum-sdk (+ nexum-runtime). The security-relevant behaviour is correct today; these harden fragile invariants.

1. Bound block_on's poll loop (robustness)

nexum-sdk/src/chain/provider.rs's block_on loops unbounded, relying on the invariant that a host call always resolves in one poll (HostTransport::call returns ready(...) today). A future alloy version chaining a retry/backoff layer, a tower::Buffer, or a pubsub / multi-await path would turn this into a silent busy-spin that burns the guest's entire gas / metering budget with no diagnostic. Cap the loop and panic with a clear message ("future did not resolve synchronously after N polls") - a fail-loud break of the one-poll invariant beats a silent hang.

2. Guard ChainMethod drift (security)

nexum-sdk/src/chain/method.rs's ChainMethod is hand-duplicated byte-for-byte against nexum-runtime/src/host/component/chain.rs's ChainMethod (18 variants, same #[strum(serialize=...)] strings), with no shared source or CI check. The guest allowlist security property ("closed enum = allowlist") depends on this matching the host dispatch table exactly. Silent drift in the SDK-diverges-from-host direction could reopen the very gap this PR closes, undetected until runtime. Extract ChainMethod into one shared definition both sides depend on, or add a cross-crate test that diffs both enums' variant strings so drift fails CI loudly.

Acceptance criteria

• block_on fails loud (bounded + diagnostic) rather than spinning unbounded.
• ChainMethod has a single source of truth, or a drift-detecting test; divergence fails CI.

[mfw78]

Decision on item 2 (ChainMethod drift): take the single-source-of-truth option, not the drift test.

Concrete shape: a new shared leaf crate nexum-chain-method (modelled on nexum-status-body), holding one pub enum ChainMethod with the derive superset Debug, Clone, Copy, PartialEq, Eq, Hash, EnumString, IntoStaticStr and the as_str(self) -> &'static str method. Sole dependency strum (already in both graphs); #![no_std] since the enum needs no allocation. nexum-sdk and nexum-runtime both depend on it and drop their hand-rolled copies. Preserve the existing paths so no consumer changes: guest re-exports at chain::ChainMethod, host re-exports at host::component::ChainMethod (keeps nexum_runtime::host::component::ChainMethod resolving for the videre-host tests). Register in workspace members, version = "0.1.0".

Both consumers are L1, so the crate stays within the L1 carve (the M5 split is not a forcing function). Gate-safe: nexum-chain-method matches none of videre|intent|venue|cow, so the venue-agnostic crate-graph scan stays green; must build for wasm32-wasip2 (the guest links it), which strum supports.

Superseding the duplicate tracker I opened at #537 (now closed) - #523 remains the single tracker and also carries the block_on bound (item 1).

[mfw78]

Item 1 (bound block_on) landed in #453 at 7797a6d: the unbounded noop-waker loop is now a single poll that panics loudly on Pending (hand-rolled, std-only, dual-target), with a #[should_panic] test. Full reasoning in the PR thread. Item 2 (ChainMethod single source of truth) remains open here.

[mfw78]

Correcting the approach I proposed above for item 2. A review of #461 points at a better vehicle than a new crate, and it checks out.

nexum-world already is the single-source vocabulary crate: pub mod caps is documented as "the single source the CORE table and the runtime's capability registry emit from", and fault_labels carries the same "single source every label mirror emits from" contract. Both nexum-sdk and nexum-runtime already depend on it. So ChainMethod should fold into nexum-world alongside caps and fault_labels, not into a new nexum-chain-method leaf crate. That drops the new-crate cost entirely and makes the fix consistent with the pattern already established for capability names and fault labels.

Supersedes the nexum-chain-method shape in my earlier comment; the rest of it (derive superset, strum-only, path-preserving re-exports at chain::ChainMethod and host::component::ChainMethod, wasm32-wasip2, gate-safe) still applies, just hosted in nexum-world.

Also worth recording from that review: #461 added #[non_exhaustive] to both ChainMethod declarations independently (nexum-sdk/src/chain/method.rs and nexum-runtime/src/host/component/chain.rs), still with no shared source. That is this issue's drift risk demonstrating itself: a change intended for the type was applied twice by hand and happened to stay in sync. Verified still duplicated at the end-of-train tip.