fix(mcp): refuse the environment writes that resolve to a wrong or hidden entry

[using-system]

Summary

The MCP server's configuration validation no longer lets three environment states through:

• a custom stack whose name a built-in ends in (monitor next to azure-monitor) is refused, and the effective entry selects <environment>-<stack> only when that key parses back to the configured pair, so a pre-existing monitor declaration can no longer read the built-in's entry;
• an environment write while the resulting stack is local is refused ("the local stack takes no environment - switch the stack first, or in the same call"), and a switch to local clears the stored environment, so nothing re-arms on the next remote switch;
• local is refused as an environment value and as a key prefix, like unknown.

The implementation choices that go beyond or narrow the proposal are recorded on the issue: #656 (comment)

Tests

• Unit: new and adapted tests in tests/mcp-server/test_config.py / test_server.py (268 passed).
• Integration: integration-tests/mcp-server/test-config-surface.sh covers the refused writes, the remote-switch-with-environment acceptance and the clear-on-local-switch.
• ruff 0.16.4 check / format, apm install --target claude + apm audit (apm-cli 0.31.0) green.

Review

A separate reviewer sub-agent checked the whole branch against the issue:

• Round 1 found one blocking defect: the built-in-suffix check also fired when an already-stored custom (monitor) split the built-in, so every new custom declaration was refused with a message blaming the wrong name. Fixed in 2403386 (only the name being declared is blamed) with a regression test. It also asked for the decision record on the issue before the PR — posted.
• Round 2: no finding, ready to merge.

Harness measurement

Coarse non-regression check only (test-plugin-harnessing, copilot / openai/gpt-5.6-luna / medium, /odd-observe drive mission on the local llms-benchmark stack, ABBA). The runtime change lives in the MCP server, which the lab runs from PyPI 1.13.1 on both sides; the only measured difference is six lines of backend-configuration/SKILL.md.

Side (mean of 2)	Turns	Commands	Tokens in	Tokens out	Wall	Preflight	Observation
main	43	39.5	3.07 M	14.3 k	415 s	71 s	216 s
branch	40	45.5	3.01 M	14.4 k	366 s	53 s	186 s

Main's spread today over six samples: turns 35–45, commands 30–51, tokens in 2.66–3.41 M, wall 310–431 s. Both branch samples sit inside it (1 premium request each, 0 authored scripts, 0 resets, 0 --help). No measurable change expected and none measured.

Closes #656

🤖 Generated with Claude Code