feat(runtime): config-mapped poison policy, retire the test setter

[mfw78]

Summary

Closes #197; part of the epic #80 train, merges in stack order.
Poison policy thresholds move from hardcoded constants into EngineConfig, and the cfg(test) back-door setter that let tests bypass the resolver is retired.

Changes

ModuleLimits gains a poison: PoisonLimitsSection field (max_failures, window_secs, both optional, #[serde(default)]) alongside a poison() resolver that follows the same saturating pattern as logs(), defaulting to the existing POISON_MAX_FAILURES / POISON_WINDOW constants when the section is absent.
window_secs uses seconds rather than milliseconds since a ten-minute quarantine window has no need for sub-second precision, unlike the http timeout knobs.
Both max_failures and window_secs saturate to a minimum of one, matching the degenerate-zero handling already used for [limits.logs], since a zero threshold or zero window would quarantine on the very first trap.
Both supervisor::boot and supervisor::boot_single now read poison_policy via limits.poison() instead of PoisonPolicy::default().
The #[cfg(test)] with_poison_policy setter is deleted; the two integration tests that relied on it (poison_pill_quarantines_module_after_threshold, multi_chain_poisoned_module_does_not_affect_other_chains) now inject tight thresholds through ModuleLimits / EngineConfig instead.
engine.example.toml gains a commented [limits.poison] block documenting the two knobs.

Test plan

• cargo fmt --all --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-features (389 passed, 0 failed), including both e2e poison tests executing against real in-tree wasms and the three new engine_config::tests::poison_limits_* cases (default, override, zero-saturation)

Notes for reviewer

The resolver returns the runtime's own PoisonPolicy type rather than a parallel engine_config-local struct, since the poison threshold already has one canonical home in runtime::poison_policy and a second copy would just be duplication; the import is same-crate and introduces no cycle.
Production behaviour is unchanged: with no [limits.poison] section present, poison() resolves to the same constants PoisonPolicy::default() used before, so both boot paths dispatch identically to main.
One pre-existing doc nit (an off-by-one "more than max_failures" wording in the operator-facing example, inherited rather than introduced here) is left as a follow-up rather than folded into this change.

AI Assistance: Claude Code (Opus 4.8 implementation and adversarial review, Sonnet 5 PR authoring) used for the full change.