Feature: cap message history in prompt at 30 recent messages

[Rustam-Z]

Summary

The harness currently injects recent Telegram message history into every CC prompt. Cap that window at 30 most recent messages per chat to keep prompts short and predictable.

Why

• Token cost. Long histories balloon the prompt — every turn pays for context the model rarely needs past the last ~30 messages.
• Latency. Shorter prompts → faster first-token, especially on -p mode (relevant until Feature: switch CC invocation from claude -p to claude --resume #42 lands).
• Cache hit rate. A bounded history is easier to keep stable across turns; an unbounded one rotates entries in/out and busts the cache.
• Quality. Long, irrelevant tails dilute attention. The recent 30 are almost always what matters; older context is better surfaced via memory / query_db on demand when actually relevant.

Behavior

• Per chat (DM or group), include only the last 30 messages in the inbound prompt.
• Reply chains (<reply_chain> blocks already injected for direct replies) still resolve outside the 30-message window — they're explicit lookups, not the rolling history.
• Older messages remain in the DB; nothing is deleted. The model can still query_db for them when needed.
• Configurable knob (e.g. MESSAGE_HISTORY_LIMIT=30 env var) so the number can be tuned without code changes.

Edge cases

• Group chats with high traffic — 30 messages may cover only a few minutes. Acceptable: model has memory + query_db for digging deeper when asked.
• DMs — 30 messages is plenty for context continuity.
• Multi-chat interleaving — limit is per chat, not global, so a busy group doesn't crowd out a DM's history.

Suggested implementation

• In the prompt-assembly path: change the message-history query from LIMIT 100 (or whatever current) to LIMIT 30, ordered by created_at DESC, then re-sort ASC for prompt insertion.
• Surface the active limit in /health for ops visibility.

Severity

Feature, low risk. Easy to ship, immediate prompt-size win.