feat: seed root trajectory id + drop vestigial open action

[ishandhanani]

Stacked on #5 (program-close); base will retarget to main once #5 lands.

Related PRs (radix-native sessions):

• feat: session radix cache sgl-project/sglang#27058 — SGLang --enable-session-radix-cache (tag + release_session, per-node holder set).
• feat(sglang): derive radix sessions from trajectory ids dynamo#10214 — Dynamo SGLang backend wiring (implicit-open, close-only session_control).
• This PR (pi-dynamo-provider) — the framework side: emits agent_context + session_control, and now seeds the root trajectory so parent_trajectory_id flows.

Verified end-to-end with all three together (GLM-4.7-Flash TP2, --router-mode kv): parent_trajectory_id flows from a real spawned subagent, no action:open, and release_session frees on close.

What

Two radix-native subagent-session lifecycle fixes:

1. Root trajectory seed. applySubagentBridge only populates parent_trajectory_id when a child inherits a non-empty DYN_AGENT_TRAJECTORY_ID. The root never set one, so the first generation of pi-subagents had nothing to inherit — the bridge no-opped and every agent showed up as its own flat top-level session (session_id == trajectory_id, no parent). This is exactly why a captured 12×4 subagent run had zero parent_trajectory_id despite the bridge being wired. seedRootTrajectory() seeds it at the root (trace on, not a PI_SUBAGENT_CHILD, id unset; prefers DYN_AGENT_SESSION_ID) so spawned subagents now carry a parent.

2. Drop the vestigial open action. The radix-native backend is close-only / implicit-open (the tag is the only session state); the first turn's action:"open" was already ignored by the worker. controlForTurn now emits a bare session_id each turn; tagging and sticky routing are unchanged (the router activates on any session_control), close is unchanged.

Also fixes pre-existing tsc errors in program-close.test.ts (optional handler invocation under noUncheckedIndexedAccess) so npm run check is green.

Test plan

• npm run check — clean (was red on the base due to the program-close.test.ts errors above)
• npm test — 40 passing, incl. new seedRootTrajectory cases (root seeds → child resolves a parent; uses DYN_AGENT_SESSION_ID; no-ops for non-root/trace-off/preset) and updated no-open controlForTurn assertions
• E2E: full stack live (Dynamo #10214 + SGLang #27058 radix worker, GLM-4.7-Flash TP2 + this provider). Root (no DYN_AGENT_TRAJECTORY_ID) → seedRootTrajectory fires → real spawned child (pi-subagents env contract) → trace records the child with parent_trajectory_id = root's id and trajectory_id=run:researcher:0 (was null in the captured 12×4 run). controlForTurn emitted a bare session_id (no action:open). KV release also verified live in --router-mode kv: the sticky KV router dispatches the deferred close_session on inline action:"close" -> worker -> release_session (freed 2 nodes on a real-prompt session; the provider's own chat-path close also dispatches, freed 0 only when the session is sub-page). In --router-mode round-robin the sticky lifecycle doesn't run, so no close dispatch — a router-mode requirement, not a regression (open-drop didn't touch close).