Two robustness fixes available: unary RPC replay correlation + goal-round-driver early disarm (fork with patches)

[mmx998]

While running deepseek-harness locally (Windows 11, v0.1.7-rc.2), we hit two issues and prepared fixes. Since external PRs are not accepted, the patches live on a public fork — happy to hand them over if useful:

• https://github.com/mmx998/deepseek-harness/tree/fix/unary-rpc-replay-correlation-20260928
• https://github.com/mmx998/deepseek-harness/tree/fix/goal-arm-until-turn-close-20260928
• https://github.com/mmx998/deepseek-harness/tree/test/projection-cache-interval-checkpoints-20260928

1. Unary RPC correlation breaks under response replay (client/connection)

An instrumenting layer between the page and the host occasionally hands a stale but well-formed response envelope to a new request, so strict rpcId matching fails: startup inventory/manifest syncs error out with rpcId mismatch. Fix: on mismatch, retry once with a fresh rpcId (these calls are idempotent); a second mismatch still surfaces. Both branches are rebased onto current master (21638c5) and follow the repo's commit conventions.

2. goal-round-driver disarms the goal while the failing turn is still open (goal/goal-round-driver)

A failing turn released the goal's armed state before the turn closed, breaking error-then-continue goal driving. Fix: keep the goal armed until the failing turn actually closes.

Testing

• pnpm exec vitest run over the owning specs: 148 tests pass across binary-rpc.host.spec.ts, client-apply.client.spec.ts, goal-round-driver.spec.ts, cache.spec.ts (includes new replay/retry coverage and armed-state coverage).
• Full host face typecheck passes after rebasing onto 21638c5.

If the team prefers the patches as a gist, email, or another format, tell me and I'll provide them.