Preserve run context for detached sub-runs - #122

Summary

• replace worker-side run-async top-level dispatch with a server-owned sub-run primitive
• recover the launching job through its short-lived nonce and carry parent stack + parent request, trace identity, and the complete secret store
• keep stack and secrets out of the runner payload and worker filesystem; the child receives only its own reader-matched grants
• retain detached execution and durable async conversation recovery

Why

Detached work should have the same execution context semantics as run-then and map-then. Serializing delegated grants into workers both duplicates the secret model and loses secrets that the parent worker itself was not authorized to read.

POST /sub-run now routes admission back to the compute thread that owns the in-flight parent. That thread validates the prepared request, clones its server-held context, and starts the child without waiting for its result. The nonce expires with the parent job.

Validation

• cargo test --workspace
• cargo fmt --all -- --check
• nix build
• CAOS_SALT=$(date --iso=s) result/bin/caos-cli run-tool test (44/44 passed)
• focused fresh-salt cargo-self and chat-offline retries after updating the EC2 test stack
• git diff --check

PR #113 will be retargeted onto this PR and will contain only conversation secret loading and request identity changes.