fix(runtime): epic mock-preset train follow-ups

[mfw78]

Summary

Closes #80; part of the epic #80 train, merges in stack order.

This PR addresses ten follow-up items raised against the mock preset and test harness work, hardening the mock chain provider, tidying shared test scaffolding, and correcting an off-by-one in the poison policy documentation.

Changes

• Added push_block_err, push_chain_log_err, close_block_stream, and close_chain_log_stream on MockChainProvider, so tests can script stream errors and stream-end conditions rather than only successful items.
• Documented the UnknownChain caveat on the mock request path: this variant conflates "chain absent" with "method not scripted" and must not be read as a chain-presence signal.
• Added pub(crate) test_utils::in_memory_logs() as the single construction site for the in-memory LogPipeline, and switched mock_components_from and both supervisor/tests.rs bundles to use it, removing duplicated construction.
• Reworded the PoisonLimitsSection rustdoc and the [limits.poison] block in engine.example.toml to state that the poison policy fires at max_failures, not one past it.
• Added a commented [limits.logs] example block to engine.example.toml, matching the resolved defaults of 262144 and 16.
• Added harness_dispatches_chain_logs, a test that drives a kind = "chain-log" module through the harness end to end, locking in the chain-log injection leg.
• Documented MockTypes as a type-level marker that is never constructed, in place of deriving Clone/Default that no call site needs.

Test plan

• cargo fmt --all --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-features (all suites pass, including the four new tests: block_stream_scripts_errors_and_end, chain_log_stream_scripts_errors_and_end, harness_dispatches_chain_logs, and the existing harness_launches... test, none skipped)
• cargo doc --workspace --no-deps with RUSTDOCFLAGS=-D warnings
• cargo clippy without test-utils enabled, confirming the mock cone stays out of a default build

Notes for reviewer

Three items from the original follow-up list are deferred rather than implemented here, each for a reason that needs more than a mechanical change:

• An optional MockRuntime preset for .runtime::<MockRuntime>() sugar needs a preset-level instance-injection path that today's config-driven presets do not have; the existing mock_components* and harness helpers already cover the in-tree need.
• Asserting guest-observed time in the clock end-to-end test needs a new guest fixture that reads wasi:clocks and writes the observed value into the mock store; the reference example module reads no clock today.
• Replacing the harness's wall-clock poll barrier with a deterministic signal needs a paused-clock drive or an event-loop readiness signal; the current 2ms/5s poll is correct but out of scope for this change.

One related item surfaced during implementation but sits outside the poison-wording item's scope: runtime/poison_policy.rs's module doc and a transition comment in supervisor.rs still read "more than max_failures" and should be aligned to "reaches max_failures" for full internal consistency with the change made here.

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