teamChat commands are lost during state summarization and never reach Agent Context

## Summary

Coach commands sent from the Player Portal through `gameState.teamChat` are received by the Agent service but are not available to the LLM during decision making.

Example commands:

```
defend
press
shoot
```

Currently, these commands do not influence player behavior.

---

## Expected Behavior

Coach instructions should be included in the player agent context:

```
GameState

↓

Context Construction

↓

Agent Context

↓

LLM Reasoning

↓

Player Action
```

The agent should be able to consider tactical instructions together with the current match state.

---

## Actual Behavior

Current execution pipeline:

```
Game Engine

↓

gameState payload

↓

agent_base.py / gateway invoke

↓

summarize_state()

↓

Agent(state_summary)

↓

LLM reasoning
```

The Agent does not receive the complete raw `gameState`.

It only receives:

```
state_summary
```

Only fields extracted by `summarize_state()` are visible to the LLM.

Current field mapping:

```
gameState

├── ball ✅ extracted
├── players ✅ extracted
├── score ✅ extracted
├── gameTime ✅ extracted
├── playMode ✅ extracted
└── teamChat ❌ not extracted
```

---

## Runtime Evidence

Logs captured before and after `summarize_state()` show that `teamChat` exists in the raw game state but disappears during context construction.

### Before summarize_state()

Input `gameState`:

```json
{
"teamChat": [
{
"senderTeamCode": "home",
"message": "defensive",
"timestamp": 0
}
]
}
```

The coach instruction exists in the incoming game state.

---

### After summarize_state()

Generated agent context:

```
Time: 11s
Score: 1-0
Team: 0 (HOME)
PlayMode: RESETTING
Ball: (7.0, 0.4) held by free
```

The `teamChat` field is missing.

This confirms the issue occurs during:

```
gameState -> Agent Context
```

conversion.

---

As a result, the coach command is not included in the LLM context during decision making.

Start in agent_base.py at summarize_state() and trace the gateway invoke path that constructs Agent(state_summary). Compare the raw gameState mapping with the generated context, focusing on teamChat. Done means coach commands are present in Agent Context and available during LLM decision making.

Written by the indexing model from the issue text.