MCP budget: default get_query_store_regressions under the response budget (#4198)

[erikdarlingdata]

Part of #4198.

Why

get_query_store_regressions had no response-size budget. McpResponseBudget.cs flagged it as the worst #4198 offender: 211 KB at default arguments on a busy production store, over 6x the 32 KB McpResponseBudget.DefaultBytes target. Two things drove that size. query_text is an unbounded Query Store text sample, one per row, at up to 50 rows by default. The ten numeric fields (durations, CPU, reads, regression percents) come straight off AVG() over microsecond integers. They serialize with a long, meaningless decimal tail instead of a display-rounded one, and that weight repeats across ten fields on every row.

What changes

Darling (DarlingMcpQueryStoreRegressionTools.cs) and Lite (McpQueryTools.cs) both change the same way:

• query_text is a 240-character preview by default (the same length get_store_query_stats already uses
  for its own full_text), with a new query_text_truncated flag. A new full_text argument (default
  false) opts back into the whole text, the MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198 precedent from get_store_query_stats.
• The default limit drops from 50 to 30. This is sized from the measured bytes per row so a default call
  stays well under budget instead of sitting right at its edge.
• The ten numeric fields are rounded to 2 decimal places at the JSON edge only. The reader's SQL and the
  desktop viewer's own read are untouched. Every existing pinned test uses whole-number fixtures that round
  trip unchanged (verified below).
• Web viewer: the /api/read row for get_query_store_regressions now passes full_text: QueryBool(c, "full_text", true) in the dispatch table, and the catalog row gets PBool("full_text", true). That keeps
  the viewer showing today's full query text (it renders it via codeDisclosure on the Query Store
  Regressions tab). No wwwroot/js change was needed: server-tabs.js's call never sent full_text
  explicitly, so the row's own default carries it. A new no-rig test, QueryStoreRegressionsWebDefaultTests.cs,
  pins this.
• FieldPreviewCutKeys (census) gets a query_text_truncated entry, added alongside the deadlock-detail
  lane's own entry. Both merged cleanly from origin/dev.
• McpToolsListBudget pins are updated on both products. TotalCeilingBytes is raised by exactly the
  measured growth (+214 Darling, +203 Lite), stacked on top of the concurrent deadlock-detail lane's own raise
  after merging origin/dev.

This PR does not touch DarlingQueryStoreRegressionReader's SQL, or Lite's equivalent reader, per the #4198
note for this tool. It only changes how the tool projects the read into JSON.

Test plan

Rig: reused the stopped rig-ta Postgres rig (port 55976, already UTC), started in the background, confirmed
"ready to accept connections", then dropped and recreated darlingtest and probe.

• New live test DarlingMcpQueryStoreRegressionsBudgetLiveTests (Darling): plants 40 regressed queries
  with ~4,091-character query text each. That fixture is the same order of magnitude as the documented
  211 KB at 50 rows before this fix. After the fix: 25,527 bytes for the default call (30 rows
  returned, truncated: true), under the 32,768-byte budget. full_text: true returns the whole text and
  no query_text_truncated.
• New Lite test QueryStoreRegressionsBudgetTests (Lite.Tests, no rig): same fixture shape, same
  assertions. It passes.
• DarlingQueryStoreRegressionsSurfaceAndSqlTests.ParamContract_AllOptional_MatchesLite updated for the
  new full_text parameter.
• McpPayloadContractCensusTests: my FieldPreviewCutKeys entry plus the deadlock-detail lane's, both
  present after the merge.
• McpToolsListBudgetTests (both products): pins and TotalCeilingBytes updated to the measured totals.
• Targeted run after merging origin/dev: Darling.Tests, the classes this PR touches plus
  McpZeroIsAMeasurementTests. 106 passed, 0 failed. Lite.Tests, the classes this PR touches.
  9 passed, 0 failed.
• Full Darling.Tests suite: not run. Every class this PR's diff touches ran targeted and green (listed
  above), but the full suite must run before arming auto-merge.
• Full Lite.Tests suite: not run, same reason.

Rig stopped (pg_ctl stop -m fast) after the last test run.

CHANGELOG entry

SECTION: Fixed
ENTRY:

• get_query_store_regressions default calls stayed under the MCP response budget ([MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198]). The
  worst-offending read measured 211 KB at default arguments on a busy store. It now previews query text to 240
  characters with a full_text opt-in, rounds its numeric fields, and defaults to 30 rows instead of 50. The
  web viewer is unaffected: it keeps the full query text it always showed.
  REF:
  [MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198]: MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198

Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3