A budget pin over every MCP read tool (part of #4198)

[erikdarlingdata]

Part of #4198.

Why

#4198 measured every MCP read tool at default arguments on two busy production stores. It found 12 tools over 50 KB and 3 over 100 KB. The worst, get_query_store_regressions, returned 211 KB. The ruling: at default arguments, every read tool stays under one shared 32 KB budget, McpResponseBudget.DefaultBytes. The per-tool PRs fixed each offender's defaults.

This PR is the test that holds that line. It measures every MCP read tool on a seeded store and fails when a tool goes over the budget without a row on file.

What changes

The diff is one new test file, Darling/Darling.Tests/McpReadToolBudgetLiveTests.cs. It uses the McpResponseBudget.DefaultBytes constant that #4205 added.

• It finds every [McpServerTool] method in the Darling service assembly by reflection. A new tool is measured with no change to the test.
• It skips analyze_*, compare_* and audit_config by name. It skips the 16 write tools through an exact list, because no name pattern catches them all. A new write tool must be added to that list, or the test calls it.
• It fills each tool's parameters the same way: the DI services, server_name set to the seeded server, and every other optional parameter at its default.
• It seeds one busy server with 12 tenant databases. The seed holds regressed Query Store pairs (about 3 KB of query text each), 30 deadlock graphs and 30 blocked-process reports. It also holds collection-log history for 20 collectors, plan-correction rows and a page of long query completions. 50 more servers, one collection each, give get_fleet_overview a fleet-sized answer.
• It measures each reply's UTF-8 bytes against the budget. A tool listed in ExemptOffenders must still be over the budget. When a fix brings a listed tool under, the test fails until its row is deleted.

The exempt list

The list began with seven tools. The per-tool PRs fixed five of them: get_query_store_regressions (#4264), get_collection_log (#4265), get_blocking (#4267), describe_custom_view_catalog (#4272) and get_query_store_top (#4273). get_fleet_overview already fit by the time this PR merged dev.

One row is left:

Tool	Bytes	Why it is still over
get_collection_health	35,145	#4268 compacts only healthy collectors with nothing to report. This fixture's collectors are not healthy, so nothing compacts.

It stays open under #4198. CI run 36142467742 measured the fixed tools: describe_custom_view_catalog 30,215 bytes, get_query_store_clutter 26,353, get_query_store_regressions 24,510, get_collection_log 21,929, get_query_store_top 16,911 and get_fleet_overview 2,087.

Known gaps

Some tools pass without a real test:

• A required parameter with no usable default. The test passes a placeholder, so the tool answers empty or invalid. These are get_custom_alert_rule, get_custom_view, get_perfmon_trend, get_plan_xml, get_query_trend, get_wait_trend, run_custom_view_panel, validate_custom_alert_rule and validate_custom_view.
• Seeded, but not enough to reach the budget. get_query_heatmap (395 bytes) needs many time buckets, and the fixture has two. get_deadlock_detail measured 19,261 bytes and get_plan_corrections 23,125.
• Not seeded. get_analysis_findings, get_index_usage, get_object_locking, get_active_queries, and every PostgreSQL-target read (get_pg_*).

Left for #4198

• get_collection_health on a server whose collectors are not healthy (the row above).
• The Lite version of this test: the same check over Lite's MCP tools against the shared DuckDB fixture. This PR does not include it.

Test plan

• GitHub Actions CI on 180c2cdc: build, Darling PostgreSQL tests and Lite tests all pass.
• The run before it, on 08e6e818, emptied the exempt list. The test failed there on get_collection_health at 35,145 bytes, so 180c2cdc put that row back. The same run's other failure, PgTargetBlockingTests, is the known flake CI flake: PgTarget anomaly/blocking worst-tile assertions fail on first attempts unrelated to the change #4274.
• The lane's first local run: McpReadToolBudgetLiveTests and DocCommentHygiene against a local rig, 78 tests, 0 failed.

CHANGELOG entry

None. This PR changes tests only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FVjn4PBJN71NQXdFo6ZxNQ