MCP budget: add get_collection_log per-server response budget and error_message preview (#4198)

[erikdarlingdata]

Part of #4198.

Why

get_collection_log's one-server form had no response-size budget. At the old 200-row
default it measured 90,514 bytes on a seeded SQL Server-target store and 89,593 bytes on
a seeded PostgreSQL-target store. Both are about 2.7x the shared 32 KB
McpResponseBudget.DefaultBytes target #4198 set. The fleet-wide form (server_name
omitted) was already sized in #4205. This is the per-server form's own pass.

What changes

• New McpResponseBudget.CollectionLogPerServerDefaultLimit (58), shared by Darling and
  Lite. Sized against the wider of the two target-kind row shapes. A SQL Server target's
  query_store rows carry the plan_fetch/text_fetch deferred-fetch split that a
  PostgreSQL target's collectors never populate. That makes the SQL Server shape
  consistently the heavier one.
• error_message is the tool's one unbounded text field. It is capped at 4000 characters
  at write time but was never previewed at read time. It is now a 500-character preview by
  default, with a new error_message_truncated flag marking a cut. Pass full_text (the
  same shape get_store_query_stats already uses) to get the whole field back. This
  guards cases a lower row limit alone does not. An unhealthy server's default
  (unfiltered) page is mostly ERROR rows near the 4000-character ceiling.
• Both changes are mirrored field-for-field in the Lite twin
  (Lite/Mcp/McpHealthTools.cs), including the tool-guide prose.
• The web viewer's /api/read row for get_collection_log (DarlingWebEndpoints.cs)
  keeps today's behavior. It already passed limit: 200 (unaffected by the new lower MCP
  default) and now also passes full_text: true. An operator reading the Collection Log
  grid still gets the whole error_message, not a preview. The two
  wwwroot/js/pages/server-tabs.js panels (Collection Log tables) call this row with
  their own explicit limit: 200 and need no change. Pinned by a new no-rig source-text
  test (CollectionLogWebRowPreservesFullTextPinTests) so a future edit to that row
  cannot drop either argument silently.
• Added FieldPreviewCutKeys to McpPayloadContractCensusTests (origin/dev did not have
  it when this lane started) and rostered error_message_truncated there. Merged with
  the two entries 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 lanes TB and the get_plan_corrections lane added concurrently.
• McpToolsListBudget pins and TotalCeilingBytes updated on both products for the new
  full_text parameter. Merged with the concurrent deltas from the same wave (see the
  change-log comments above each constant).

The fleet-wide form (server_name omitted) is untouched, as #4205 already sized it.

Test plan

Darling, against a rig seeded on both target kinds (200 rows/server, mixing server-scoped
and enumerated collectors, 10% carrying a moderate error_message):

• New McpCollectionLogPerServerResponseBudgetLivePostgresTests: before the fix, SQL
  Server target measured 90,514 bytes and PostgreSQL target 89,593 bytes (both over
  budget). After the fix: 28,340 bytes and 28,057 bytes (both under the 32,768-byte
  budget).
• After the merge with origin/dev, these all passed: McpPayloadContractCensusTests,
  McpToolsListBudgetTests, PgCappedReadSurfaceTests,
  CollectionLogWebRowPreservesFullTextPinTests (new), DarlingCollectionLogReadTests,
  McpFleetResponseBudgetLivePostgresTests, McpFilterSemanticsLivePostgresTests,
  AsOfWindowAnchorTests. 73 tests in the first group's run, 15 in the second.

Lite (no rig, local DuckDB):

• New CollectionLogPerServerResponseBudgetToolTests: default call on 200 seeded rows
  stays under budget.

• McpToolsListBudgetTests, CollectionLogFleetFormToolTests, CollectionLogToolTests:
  all green.

• Full Darling.Tests suite (targeted classes above were run individually. The full
  suite was not run in this lane due to context budget).

• Full Lite.Tests suite (same reason).

Both builds are 0 Warning(s), 0 Error(s).

CHANGELOG entry

SECTION: Fixed
ENTRY:

• MCP get_collection_log stays under the response-size budget by default ([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]) -
  At the old default, a per-server call returned close to 90 KB. That exceeds the shared
  32 KB budget other read tools already honor. The default row limit is now sized to the
  tool's own row shape. The error message field is a 500-character preview by default
  (pass full_text to get it whole). The web viewer keeps its existing behavior.
  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_01FVjn4PBJN71NQXdFo6ZxNQ