fix(mcp): prevent OOM from long-lived stream retention

[JayZeeDesign]

What this does

Fixes production OOM caused by long-lived MCP stream retention. The incident pattern: GET /mcp/ connections from Cursor, claude-code, and opencode hold SSE streams open for minutes to hours, accumulating memory (~10 MiB worst-case per connection) until RSS spikes from ~1 GiB to ~3+ GiB and OOM kills the pod at 4 GiB.

Adds MCPStreamLimitsMiddleware that enforces connection lifecycle limits:

1. Per-instance concurrent connection limit (default 200) — reject new connections at 503 when at capacity
2. Per-org concurrent connection limit (default 20) — fair sharing across teams, prevents one team from monopolizing connections
3. Max connection lifetime (default 30 min) — force reconnection after N seconds; buffers are freed, clients auto-reconnect
4. Idle timeout (default 5 min) — close connections with no activity (no receive/send)
5. Prompt cleanup on disconnect — ensures buffers are freed when client disconnects (499)
6. Structured logging with metrics — periodic snapshots of connection count, age distribution, rejection counts

Configuration

New environment variables in config.py:

Variable	Default	Purpose
TREG_MCP_MAX_CONNECTIONS	200	Max concurrent MCP connections per instance
TREG_MCP_MAX_CONNECTIONS_PER_ORG	20	Max concurrent connections per org (fair sharing)
TREG_MCP_MAX_LIFETIME_S	1800	Force reconnection after N seconds
TREG_MCP_IDLE_TIMEOUT_S	300	Close connections idle for N seconds

Why these defaults

• 200 connections × ~10 MiB = 2 GiB headroom on a 4 GiB instance with ~1 GiB baseline
• 30 min lifetime is long enough for most tool sessions, short enough to prevent unbounded accumulation
• 5 min idle is conservative; SSE streams typically have keepalive traffic
• Per-org limit ensures no single team can cause OOM

How it was tested

1. Unit tests — 23 tests covering all limit types, timeout scenarios, disconnect cleanup
2. Integration tests — verified existing MCP test suite (136 tests) still passes
3. Import linter — all 14 contracts pass
4. Manual verification — middleware correctly wraps MCP transport

# All new tests pass
uv run pytest tests/test_mcp_limits.py -v  # 23 passed

# Existing MCP tests unaffected  
uv run pytest tests/test_mcp.py tests/test_mcp_directory.py -q  # 113 passed

# Import rules maintained
uv run lint-imports  # 14 kept, 0 broken

Verification in production

After deploy:

1. Check logs for mcp_stream_metrics entries — should show reasonable connection counts
2. Monitor RSS — should plateau instead of growing unbounded under MCP load
3. Watch for mcp_connection_rejected logs — indicates limits are working
4. Confirm no user-visible impact — clients should auto-reconnect seamlessly

Rollback

Set very high limits to effectively disable:

TREG_MCP_MAX_CONNECTIONS=10000
TREG_MCP_MAX_LIFETIME_S=86400
TREG_MCP_IDLE_TIMEOUT_S=86400

Checklist

• uv run --with pytest-xdist pytest -n auto -q passes locally (ran subset; full suite in CI)
• Added or updated tests for the change (23 new tests in test_mcp_limits.py)
• Updated the relevant docs/context/ fragment (no existing MCP fragment)
• No secrets in the diff (keys, tokens, .env values)