feat(swarm): improve trail-hunter threat awareness

[christopherjnelson]

Summary

• Include current-tick observable disinfections in bounded worker pressure history before reflex decisions, without committing player events early or duplicating them.
• Give Agent Zero each surviving worker's local pressure and semantic direction/distance, plus bounded recent capture facts. Explain permanent trail-hunter capture and abandoned territory in the planner instruction.
• Raise worker replan signals at 0.50 under rising/high pressure while retaining 0.80 under low pressure. Ignore stale requests from workers captured on the current player advance.

Evidence and scope

A deterministic service test exercises a real hunter advance: the worker infects and leaves a cell, the hunter disinfects it within one cell on tick 3, Jev-like probability 0.60 produces a worker request, and Zero sees worker-request on tick 4. The test also verifies low-pressure 0.60 decisions stay silent and a later capture removes the worker and its stale request from Zero's observation.

The hunter engine, mission set, directive completion, Jev policy, legacy cognition, provider accounting, and raw-provider export boundaries are unchanged. No live provider calls were made.

Validation

• pnpm validate — passed locally.
• pnpm test:e2e — repository owner confirmed green after stopping occupied development ports. The same two Playwright browser tests also passed in an isolated copy on alternate ports.
• Focused game-api, shared, and agent-runtime tests — passed.
• Fresh reviewer challenge addressed the captured-worker stale replan request.

Follow-up

After merge, manually rerun the same 50-tick Zero/Jev + trail-hunter scenario before changing target selection, mission semantics, or hunter balance.