feat(agent): consume similar_trajectories in plan construction — retrieved high-reward trajectories are fetched then ignored #10581

similar_trajectories are retrieved with reward filtering, then ignored by plan construction

Part of umbrella #10542 · concrete sub-slice of #10545 · follow-up to the trajectory store (GH#7357).

Research origin: Ornith-1.0 comparison — high-reward past trajectories should bias the next plan. AutoBot already retrieves them (paying the vector-search cost) but never feeds them into planning.

The gap (verified)

WorkflowPlanner fetches high-reward similar past solutions before planning:

• autobot-backend/orchestration/workflow_planner.py:288-293 — find_similar_trajectories(user_request, top_k=5, min_reward=0.7) → stored in context["similar_trajectories"].

But:

• The method docstring at :99 states plainly: "Step generation always [builds from scratch]"; callers may only "inspect via context['similar_trajectories']."
• Grep confirms zero consumers read context["similar_trajectories"] (only the populate site and the docstring reference it).

Net: we run a ChromaDB query + embedding on every plan, filter to reward ≥ 0.7, and then discard the result. The reward signal stored in the trajectory store influences nothing.

What to build

• Inject the retrieved trajectories' action_sequence / strategy into the planning prompt as few-shot priors ("a similar high-reward task was solved via …"), so the planner can reuse proven decompositions instead of always re-deriving.
• Keep it advisory-but-used: the planner may deviate, but the trajectory must demonstrably reach the prompt.
• Non-fatal path already exists (try/except at :299) — preserve it.

Acceptance criteria

• When a high-reward similar trajectory exists, its decomposition/strategy is present in the planner's prompt (assert on the rendered prompt in a test).
• With no similar trajectory (or all below min_reward), behavior is unchanged.
• No change to the retrieval contract in memory/trajectory_store.py.

Entry points

autobot-backend/orchestration/workflow_planner.py (populate + step generation) · autobot-backend/prompt_manager.py · autobot-backend/memory/trajectory_store.py (read-only)