fork of prime agent with changes i want but are not upstreamable probqbly (and some fixes)

fix(coding-agent): pin the child cwd, harvest silent children, backpressure fan-in (R4, R11, R10, R9) master

R4 - children defaulted to whatever the host happened to pass, so recursive searches walked multi-TB caches and froze kernels for 21 and 92 minutes (01a09d7f, 01a09bf6, 01a09351). `CreateRlmSubagentRuntimeOptions` now carries an explicit `cwd`, the parent always sets it to its own session directory, and both runtime hosts and the inline path use it instead of reaching into the parent session. Proof: test/suite/subagent-cwd-propagation.test.ts. R11 - 11 consecutive children exited without replying (01a09782) while parents waited on fan-in that never came. A child that finishes with zero `agent_message.send` calls now hands its last assistant text to the parent as `[auto-harvested fallback from <name>]`, tagged as unreviewed, with an empty child still getting the plain no-reply notice. The reply obligation also moved into the task envelope and the child doctrine, where a child cannot miss it. Proof: test/suite/subagent-reply-harvesting.test.ts. R10 - fan-in hit a hard 20-pending ceiling and threw at the exact moment results arrived (01a09d35). The ceiling is now soft (200) and admission never fails on it: the message buffers in memory, and the daemon yields up to 30s for a saturated recipient to drain before delivering anyway. The throwing `assertAgentMessageQueueCapacity` helper is gone; the two tests that asserted the old fatal cap now assert the resilient behaviour. Proof: test/suite/agent-message-backpressure.test.ts (35 concurrent senders, zero rejections). R9 (host half) - the worker forwarded PATH alone to a session it created, so `cargo`/`nix` were missing under a GUI- or launchd-started daemon (01a09746 exit 127). The worker now forwards the launching shell's environment, which the supervisor already seeded it with.