get_collection_health's default fits the MCP response budget (#4198)

[erikdarlingdata]

Closes #4198.

Why

get_collection_health was the last tool on both SKUs' ExemptOffenders lists (#4198's per-tool budget
backlog). #4268 already compacted HEALTHY collectors with nothing to report to a 7-field shape, but a row that
needs a look (failing, stale, stopped, erroring, denied, regressed, or an unexplained zero) still kept the full
~30-field shape by default, and on both fixtures that alone was over McpResponseBudget.DefaultBytes (32,768 B):
Darling 35,145 B (20 rows), Lite 43,115 B (42 rows). Lane M1 measured this in detail and designed the fix
(issue comments) but ran out of context before writing code; this lane implements it.

Lane M4b follow-up (2026-09-25): the leaner shape this PR shipped named errors and zero-rows as reasons a
row needs a look, but not the other three the predicate itself checks: session-missing runs, abandoned cycles,
or a permission denial cleared by a later success. A HEALTHY row whose only issue was one of those three carried
partial_detail: true with nothing in the shape saying why. Fixed below; see "What changes" and "Measured
bytes" for the updated field list and numbers.

What changes

A row that fails IsCollectionHealthCompactEligible now gets a new leaner shape instead of the full one,
on both SKUs, field-for-field identical:

collector, status, partial_detail, total_runs, errors, session_missing, abandoned, rows_stored, last_success, last_error, last_error_truncated, last_error_at, last_denied_at, denied_since_last_success, regressed_from_productive, output_finding, output_finding_truncated

• partial_detail: true is the row marker — deliberately not compact, whose meaning is "healthy, nothing
  to report." CollectionHealthPayloadBudgetLiveTests (Darling) and CollectionHealthPayloadBudgetToolTests
  (Lite) assert a row that needs a look never carries compact, so a caller scanning for compact != true
  still never skips it.
• last_error previews to the existing ErrorMessagePreviewLength (500 chars, shared with get_collection_log);
  output_finding (a template sentence from FormatOutputFinding, not raw free text, and the single dominant
  cost on Darling's fixture where every synthetic row hit the "needs a look" branch) gets its own shorter
  OutputFindingPreviewLength (200 chars). Both _truncated flags are always emitted (true/false), matching
  get_collection_log's error_message_truncated — that flag is not conditionally omitted, so lever (a) in the
  brief (omit the flag unless true) does not apply; McpHelpers.JsonOptions writes nulls and the row is an
  anonymous type, so a conditionally-omitted property isn't structurally available here anyway.
• session_missing, abandoned and last_denied_at are now in the leaner shape (lane M4b, 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
  follow-up), same names and formats FullCollectionHealthRow uses, always emitted rather than conditional on
  being the reason a given row is partial. ExtensionMissingCount is also in
  IsCollectionHealthCompactEligible's predicate, but it surfaces under no field name anywhere in the full
  ~30-field row either — there is nothing for the leaner shape to mirror — so nothing was added for it.
  Remeasured after adding the three fields: still comfortably under budget on all four fixtures (numbers below).
• full_detail=true is unchanged: every field on every row, full stop.
• collector_detail_note now names both tiers (how many rows compacted, how many took the leaner shape) and the
  one restore sentence (full_detail=true) that applies to both.
• Both [Description] texts' full_detail (#4198) paragraph is corrected — the old sentence claiming every
  other row "always carries every field" is no longer true — and kept byte-identical between the two SKUs,
  confirmed by CrossSkuSurfaceSourceTests.BothSkusToolDescriptions_StayByteIdentical (7/7 pass, including after
  the M4b field-list edit). The edit lives entirely after the <<GUIDE>> marker (the tail get_tool_guide
  serves separately), so McpToolsListBudgetTests' TotalCeilingBytes on both SKUs measured unchanged
  (92,222 Lite, 174,554 Darling) — ran both, no constant edit needed.
• McpPayloadContractCensusTests (Darling) does not need new entries for the three added fields: its
  SecondBoundCutKeys / CutNoteKeys / SourceSideCutKeys / FieldPreviewCutKeys arrays census
  truncation/preview keys specifically (the *_truncated flags and their preview lengths), and
  session_missing/abandoned/last_denied_at are plain mirrored counts and timestamps, not previews — none
  of the four arrays applies to them. Confirmed by running the class: still green, no new failures asking for an
  entry.

Correction to M1's design notes: Lite already had a dedicated payload-budget test,
CollectionHealthPayloadBudgetToolTests.cs (added by #4268, before M1's session), not "no twin of Darling's
payload test" as M1's notes said. I extended the existing test with the 80% ceiling and marker assertions
instead of adding a new file.

Measured bytes

Fixture	Before this PR	After leaner shape (M1)	After M4b's 3 fields	Budget	80% ceiling
Darling McpReadToolBudgetLiveTests (mixed SUCCESS/ERROR, 20 collectors, ExemptOffenders)	35,145 B	21,027 B	22,147 B	32,768 B	26,214 B
Lite McpReadToolBudgetTests (mixed SUCCESS/ERROR, 42 collectors, ExemptOffenders)	43,115 B	28,017 B	30,369 B	32,768 B	26,214 B
Darling CollectionHealthPayloadBudgetLiveTests (realistic: mostly healthy + 6 pinned non-boring rows)	n/a (new assert)	18,418 B	19,134 B	32,768 B	26,214 B
Lite CollectionHealthPayloadBudgetToolTests (same shape)	n/a (new assert)	18,366 B	19,082 B	32,768 B	26,214 B

Both ExemptOffenders rows are still gone; both dictionaries are still empty (this PR still closes #4198, not
just advances it — the M4b fields did not put either fixture back over budget).

The realistic dedicated fixtures grew two more pinned rows each (six, not four): plan_cache_stats
(session-missing only) and tempdb_stats (abandoned cycles only, 1 in 250 runs to stay at a 0.4% abandon rate,
under the 0.5% WARNING cutoff so the row still bands HEALTHY). The third case — a permission denial cleared by a
later success — was already covered by the existing query_store_health row; it now also asserts
last_denied_at is non-null. All three new pin assertions were proven to fail against the pre-M4b shape (checked
out the parent commit's two production files, rebuilt, confirmed [FAIL] on both SKUs' dedicated tests, then
restored and rebuilt back to green) before being left in place.

One open point for the coordinator, updated: the ruling's item 6 says a default call must land at 80% of
budget "to leave room for a larger fleet." Both dedicated tests still clear the 26,214 B ceiling comfortably
(~27% headroom, down slightly from ~30%). Darling's ExemptOffenders fixture also still clears it (22,147 B).
Lite's ExemptOffenders fixture does not (30,369 B) and its margin under the hard 32,768 B budget has
thinned from 14% to 7.3% (2,399 B) after the M4b fields — worth watching if a future field gets added to
this shape, since that fixture is already the tightest of the four. Per the brief, I did not cut fields to buy
margin back; flagging the number for the coordinator instead.

Test plan

• Lite build: dotnet build Lite/PerformanceMonitorLite.csproj — 0 warnings, 0 errors.
• Darling service build: dotnet build Darling/PerformanceMonitor.Darling.Service/...csproj — 0 warnings, 0 errors.
• Lite.Tests.CollectionHealthPayloadBudgetToolTests, Lite.Tests.McpReadToolBudgetTests — pass, bytes above.
• Darling.Tests.CollectionHealthPayloadBudgetLiveTests, Darling.Tests.McpReadToolBudgetLiveTests (rig, port 55975) — pass, bytes above.
• Lite.Tests.CrossSkuSurfaceSourceTests (description byte-identical pin) — 7/7 pass.
• Lite.Tests.McpToolsListBudgetTests, Darling.Tests.McpToolsListBudgetTests — pass, totals unchanged.
• Full Lite suite (Lite.Tests.exe, no filter): 5,389 total, 0 failed.
• Full Darling suite (Darling.Tests.exe, no filter, fresh darlingtest, rig in UTC): first run found 2
  failures — DocCommentHygieneTests.NoMemberCarriesTwoStackedSummaryBlocks (an ExemptOffenders edit left
  the old <summary> block stacked above the new one instead of replacing it) and
  McpPayloadContractCensusTests.EveryCutKey_IsThePageDialect_OrClassified_OnBothSkus (the two new
  *_truncated keys needed a FieldPreviewCutKeys census entry). Both fixed, reconfirmed green (145/145)
  plus the collection-health-specific classes re-run clean. Did not re-run the full 14-minute suite a third
  time given context budget.
• Lane M4b (this update), after git merge origin/dev: rebuilt both test projects (0 warnings, 0
  errors) and ran, by class name: Darling.Tests.CollectionHealthPayloadBudgetLiveTests,
  Darling.Tests.McpPayloadContractCensusTests, Darling.Tests.CrossSkuSurfaceSourceTests,
  Darling.Tests.McpToolsListBudgetTests, Darling.Tests.McpReadToolBudgetLiveTests (76/76 in one run,
  rig port 55975), then Darling.Tests.CollectionHealthAggregateTests +
  Darling.Tests.DocCommentHygieneTests (90/90) — this last pair is the subset most likely to catch a
  problem from this lane's own new doc comments and test edits. On Lite:
  PerformanceMonitorLite.Tests.CollectionHealthLatestNoteTests, CollectionHealthPayloadBudgetToolTests,
  CollectionHealthWindowTests, CrossSkuSurfaceSourceTests, McpReadToolBudgetTests,
  McpToolsListBudgetTests (12/12 in one run). All pass; measured bytes above.
• Manual revert-and-rerun (lane M4b): checked out the parent commit's two production files over the working
  tree, rebuilt, confirmed the three new pin assertions [FAIL] on both SKUs' dedicated budget tests,
  then restored and rebuilt back to green — proves the pins actually test the new fields rather than
  passing vacuously.
• Full Lite suite / full Darling suite: not re-run by lane M4b (brief: "No full suite: CI runs it"). This
  lane's changes are three new object-initializer fields, guide-text edits, and additive test assertions in
  the same classes already exercised above; no other subsystem touched.
• Rig stopped and C:\GitHub\worktrees\rig-m4b deleted after lane M4b's runs.

CHANGELOG entry

SECTION: Fixed
ENTRY:

• get_collection_health's default reply now fits the MCP response budget ([get_collection_health's default fits the MCP response budget (#4198) #4319]) - A collector row that
  needed a look (failing, stale, stopped, erroring, denied, regressed, or an unexplained zero) used to keep
  every field by default, and on a server with several such collectors the reply could exceed the 32 KB MCP
  response budget. That row now gets a leaner shape carrying only what says a collector is wrong and since when;
  full_detail=true still returns every field on every row.
  REF:
  [get_collection_health's default fits the MCP response budget (#4198) #4319]: get_collection_health's default fits the MCP response budget (#4198) #4319

🤖 Generated with Claude Code

https://claude.ai/code/session_01FVjn4PBJN71NQXdFo6ZxNQ