refactor(runtime): route fixed session identity through SDK

[limityan]

Summary

• add a source-compatible AgentRuntime::create_session_with_id path while keeping AgentSessionCreateRequest unchanged
• use the exact-ID path for exec --session-id and CLI backend-session recreation; retain the old Core facade method as a deprecated migration shim
• reject invalid, duplicate, unsupported, or provider-mismatched identities with typed errors while preserving the existing create_session error contract
• make first persistence claim the final session directory atomically, so only its owner can clean up a failed write and a concurrent loser cannot delete the winner
• update the existing runtime and CLI architecture facts without adding a standalone design document

Product and architecture boundaries

• SDK API version remains v1: the new port method has a default NotAvailable implementation and no field was added to the public create DTO
• ordinary CLI/TUI creation, ACP, remote-connect, and agent-tool paths still request generated IDs
• no new TUI screen, command, selector field, session owner, generic persistence transaction, extension surface, or HarmonyOS behavior
• model updates, branching, usage, snapshots, persistence maintenance, and loaded-session checks remain on the existing compatibility path
• the internal first-write API is crate-private; ordinary session updates continue to use the existing save_session path

Failure and concurrency behavior

• late metadata/state write failures roll back memory, caches, storage claims, the owned session directory, and its index entry; retrying the same ID is covered
• two independent persistence managers sharing one root can race on the same ID, but exactly one can acquire the final directory and the loser cannot write or clean it
• a provider that reports the wrong ID is rejected by Runtime; Runtime cannot undo side effects already performed by a non-conforming provider
• a process crash after the directory claim can still leave a partial directory that is safely rejected on retry; crash recovery is intentionally outside this PR

Validation

• cargo +nightly test -p bitfun-runtime-ports
• cargo +nightly test -p bitfun-agent-runtime
• cargo +nightly test -p bitfun-core agent_submission_create_session_ -- --nocapture
• cargo +nightly test -p bitfun-core concurrent_first_session_persistence_keeps_the_winner
• cargo +nightly test -p bitfun-core persistent_session_creation_failure_does_not_publish_runtime_state
• cargo +nightly test -p bitfun-core compatibility_operations_have_one_core_owned_facade
• cargo +nightly test -p bitfun-cli
• cargo +nightly check --workspace
• node scripts/check-core-boundaries.mjs
• pnpm run check:repo-hygiene
• git diff --check gcwing/main

Rust validation used the coherent installed nightly toolchain because the local stable Cargo and standard-library installation are mismatched. Existing dependency deprecation and CLI dead-code warnings remain unchanged.