Close out #3898: bank the tools/list savings, drop the last issue ref

[erikdarlingdata]

Closes #3898.

Why

#3898 split every MCP tool's description into a short served head plus an on-demand reading guide. It also set a growth-only ceiling on the total tools/list payload, so a later PR cannot quietly regrow it. Every content PR that converted a tool family, and both feature PRs, have now merged to dev.

This PR is the close-out. It banks the saving those PRs already earned, by lowering both ceilings to what the payload actually measures today. It also removes the one issue reference (#3653) still riding on the wire in shared text.

What changes

Both total ceilings drop to the measured total

Product	Before	After	Cut
Darling	329,494 bytes	170,798 bytes	158,696 bytes (48.2%)
Lite	148,216 bytes	89,719 bytes	58,497 bytes (39.5%)

The "before" figures are each ceiling's value as first committed in #3898's Phase 0 (git log -S TotalCeilingBytes). Neither ceiling was ever lowered before this PR. Each was only raised once, deliberately, in the #4048 D9 round. The "after" figures are the measured totals with this PR's change applied. They come from the budget test's own failure message, read with the ceiling temporarily set to 0.

McpHelpers.WindowTruncatedDescription drops its issue reference

This sentence is appended to several tools' descriptions on both products to explain the window_truncated field. Issue references come off unless they carry a rule (D4). WIRE CHANGE: spelled truncated before #3653. becomes WIRE CHANGE: formerly named truncated. The fact stays: the field used to be called truncated. The lane's first version, spelled truncated before., no longer said before what. The PR tender reworded it at the same length. Every other word in the constant is unchanged.

That six-character shrink only reaches a tool's served (budgeted) text where the constant lands in the head. That happens on exactly one tool per product, get_query_trend, which drops from 798 to 792 bytes on both Darling and Lite. On every other tool that carries the constant, it is appended after the <<GUIDE>> marker. There it lands in the tail get_tool_guide serves on demand, and the served head does not move. Two test assertions in McpPayloadContractCensusTests.cs quoted the old literal text. Both are updated to match.

Converted tool counts, and what still serves more than 620 characters unconverted

Product	Total tools	Converted (carry <<GUIDE>>)	Unconverted
Darling	159	105	54
Lite	89	52	37

Darling tools still serving more than 620 characters unconverted:

Tool	Characters
analyze_query_store_plan	914
get_query_heatmap	799
get_table_index_sizes	794
get_pg_wait_trend	792
get_pg_wraparound_risk	792
get_query_trend	792
update_custom_view	785
get_pg_deadlock_detail	759
get_waiting_tasks	751
get_pg_lock_stats	722
get_pg_buffer_usage	684
get_object_locking	661
get_server_summary	648
get_lock_wait_trend	646
get_running_jobs	642
get_cpu_scheduler_pressure	640

Lite tools still serving more than 620 characters unconverted:

Tool	Characters
get_query_heatmap	799
get_table_index_sizes	794
get_query_trend	792
get_waiting_tasks	751
get_object_locking	661
get_server_summary	648
get_lock_wait_trend	646
get_running_jobs	642
get_cpu_scheduler_pressure	640

analyze_query_store_plan stays unconverted on purpose. It is not in Lite's list above because Lite already serves it in 210 characters. Darling serves it in 914. An identical converted head grows Lite's payload instead of shrinking Darling's, so it stays as is.

This PR converts no tool and edits no description other than WindowTruncatedDescription. There is no changelog entry: every converted tool already has one buffered from its own content PR.

PR tender verification

• Head facts fixed: 1, the coordinator's finding. Commit 0039c09d changes the notice to
  WIRE CHANGE: formerly named truncated. It is the same 25 characters as spelled truncated before., so
  neither ceiling nor the get_query_trend budget lines move. The census test's assertion and the comment below
  it quote the new text.
• The doc comment above the constant said no on-demand guide exists to carry the notice. That went stale when
  The MCP surface costs ~100k tokens of context before the first call: 157 tools, 336 KB of tools/list, 86 KB of instructions #3898 shipped get_tool_guide. It now says a split tool carries the notice in its guide tail, and an unsplit
  tool still serves it in the tool list. Today the only unsplit one is get_query_trend.
• Both ceilings are exact. With each ceiling one byte lower, ToolsList_TotalBytes_IsAtOrUnderTheCeiling failed
  on both products and reported totals of 170,798 bytes (Darling) and 89,719 bytes (Lite). Restored, it passed.
• No open PR touches an MCP file. 4053 c1: RDS/Aurora pg_log_events reads the csvlog file when csvlog is on #4135, 4053 part b1: pg_deadlocks reads csvlog when it's on #4136, 4053 part b2: pg_plan_capture reads csvlog when it's on (self-hosted) #4137 and 4053 review round 2: caller-stated csvlog edges + narrower log-format cache invalidation #4140 were checked, so none of them can push the total
  past the new ceilings.
• D9 round 1 (haiku, 3 questions on the 792-byte served text of get_query_trend): 0 misreads. All 3 answers
  were correct: the old key is gone, no limit moves the window floor, and the old name was truncated.
• Targeted tests after the fix: Lite 120 total, 0 failed. Darling 323 total, 0 failed, 8 skipped. The classes:
  *McpPayloadContractCensus*, *McpToolsListBudget*, *McpToolGuide*, *McpDescription*,
  *MeasurementContractCensus*, *McpPageContract*, *McpZeroIsAMeasurement*, *WindowTruncat* and
  *QueryTrend*.
• Changelog: the buffer entry 4141-reported.txt was missing. The tender wrote it.

Test plan

• Darling.Tests builds with 0 Warning(s), 0 Error(s).
• Lite.Tests builds with 0 Warning(s), 0 Error(s).
• Darling: McpToolsListBudgetTests, McpPayloadContractCensusTests, QueryStoreClutterTests, QueryStoreTopWindowTests (the classes a git grep for WindowTruncatedDescription finds), and every McpToolGuide* class. 280 total, 0 failed.
• Lite: McpToolsListBudgetTests and every McpToolGuide* class. 80 total, 0 failed.
• Darling full suite, once: 13,471 total, 0 failed, 638 skipped. The skips are live PostgreSQL tests. This lane runs no live rig. 1 test did not run, a pre-existing row unrelated to this change.
• Lite full suite, once: 5,299 total, 0 failed, 0 skipped, 0 not run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QQh8LczP1HGLXQYAx4oTZC