Fault-injection mock host: latency, dropped handshake, flaky transport

Summary

Extend the strict mock host with fault-injection knobs so SDK-level failure modes are reproducible in CI: injected latency, a dropped/never-answered handshake (a permanent repro for product-sdk#200), a flaky transport, and a version-skewed host. Today these are only reproducible against real hosts, by hand, after the fact.

Motivation

The test host is faithful (real signing, real SCALE, SS58-only matching) but always happy-path. The bugs that bite are timing/transport artifacts a canned mock can't reproduce - product-sdk#200 (a 0.8.4 client silently drops the handshake against a 0.8.0 host; isReady() hangs), and the desktop ChainHead churn hang. A fault layer over the real transport reproduces them deterministically.

Design

A thin layer wrapping the raw @novasamatech/host-api Provider, read on every message so faults work up front and at runtime:

createTestHostServer({
  productUrl,
  faults: { latencyMs?, dropHandshake?, dropEveryNth?, protocolVersion? },
});

Plus FAULT_SCENARIOS presets (droppedHandshake, flakyTransport, versionSkew, slowSigning, highLatency) and runtime setFaults/getFaults on the fixture. Handshake frames are recognised by reading the SCALE-compact requestId length + the MessagePayload enum index (response = 1, request = 0) - no dependency on non-exported codecs. protocolVersion rewrites only the inbound handshake request's trailing codec-id byte, so the host's built-in handler emits the real UnsupportedProtocolVersion (no fake/encoding on our side). No-fault config is a transparent pass-through.

Done (feature-complete, green locally)

• Fault layer (src/fault-provider.ts), DOM-free; bundled by esbuild + unit-testable under node.
• faults option + setFaults/getFaults through runtime + Playwright fixture; FAULT_SCENARIOS.
• dropHandshake (product-sdk#200 repro), latencyMs, dropEveryNth, protocolVersion (version skew - real UnsupportedProtocolVersion path).
• 13 unit cases + 5 E2E specs; full integration suite 61/61 green.
• Mandated docs updated (CHANGELOG, forum-post, README).

Out of scope - deferred to the @parity/truapi-host engine swap

• flakyTransport recovery (true retry): asserting the SDK recovers a dropped request needs a per-request timeout/retry in the transport and a retry-capable signer in the loop - neither exists on the @novasamatech engine (a dropped request hangs with nothing to reject on). The shipped flakyTransport proves the fault is observable (the call stalls); recovery follows [ENG].
• Literal-semver version negotiation: the wire carries a u8 codec id, not a semver. A semver-string negotiation needs the @parity/truapi-host engine.

Acceptance criteria

• faults + setFaults/getFaults documented and exported.
• droppedHandshake E2E fails the way product-sdk#200 fails.
• versionSkew E2E: handshake rejected, product does not connect.
• flakyTransport E2E: a dropped signing request is observably stalled.
• No-fault config is a transparent pass-through (existing tests unaffected).

Notes

This is migration-independent - it ships on the current @novasamatech engine and does not wait for product-sdk#186. Implementation is complete and green locally; the PR is a rebase onto current main + a version re-pick (0.10.0 already shipped without this → 0.11.0) away.

Start with src/fault-provider.ts, then inspect the runtime and Playwright fixture exports for faults, setFaults, getFaults, and FAULT_SCENARIOS. Run the 13 unit cases and 5 E2E specs, followed by the full integration suite, and verify the documented acceptance criteria and no-fault pass-through on current main.