MCP budget: finish get_analysis_findings default limit and text previews (#4198)

[erikdarlingdata]

Part of #4198.

Finishes lane TG's work on get_analysis_findings (both products), picked up from checkpoint e159afd5 after TG stopped at its 200-turn limit.

Why

get_analysis_findings had no default response-size budget. A busy production store's default call (24h, no limit) measured 66,838 bytes, well over the shared 32 KB McpResponseBudget.DefaultBytes. TG sized the fix. A default limit of 18 diagnostic chains caps the page, and findings_truncated flags a cut. Each finding's confidence_basis and its advice (investigation/remediation) preview to 160 characters by default, with *_truncated flags and a full_text opt-in. remediation_command is never previewed, because a half risk-disclosure header on a destructive change is worse than none.

What changes

• Built both test projects and fixed what did not compile against the checkpoint's own targets.
• The limit parameter's description was 225 characters, over the D2 200-char cap for a converted tool's parameter. Cut it to 122 characters and moved the detail into the tool's guide tail, on both products.
• Fixed a missing space the checkpoint's head-sentence insertion left ("whole.confidence_basis").
• Caught a second, stricter budget. Lite's full suite (McpToolGuideTests, McpToolGuideHeadsSqlCoreTests) enforces a hard 620-character served-head cap that McpToolsListBudgetTests' looser check (1000 hard, 600 target) does not. The checkpoint's head-sentence addition pushed the served head to 740 characters. That fact is not in either guide's required guardrail roster. It is now covered in the tail, so the sentence is dropped from the head entirely. The head is back to its exact original text: 610 served bytes on both products.
• Fixed the /api/read row for get_analysis_findings in DarlingWebEndpoints.cs. It had inherited the new limit: 18 / full_text: false defaults. That limited the web viewer's two "Analysis Findings" tables to 18 of a window's chains. The tables render no field the text preview cuts, so only the row-count cut was visible. It now passes Rows(c, "limit", MaxRowLimit) and QueryBool(c, "full_text", true), matching what dev always returned: every chain, full text. Added a no-rig source-scan pin in DarlingWebEndpointsTests (GetAnalysisFindingsRow_PassesTheOldViewerDefaults_EveryChainFullText) that fails if the row regresses. GetAnalysisFindings needs a live Postgres connection, so the row cannot be pinned by invoking it directly.
• Census: classified the four new payload keys in McpPayloadContractCensusTests. findings_truncated and findings_truncated_note join SecondBoundCutKeys and CutNoteKeys as a second, independent page cut beside the tool's pre-existing truncated/truncation_note. confidence_basis_truncated and advice_truncated start a new FieldPreviewCutKeys roster. It uses the same name and tuple shape #4254 uses for get_deadlock_detail, so a later merge only needs to combine array entries.
• Updated both products' McpToolsListBudgetTests pins: the two new parameter lines (limit 122, full_text 146), and each product's TotalCeilingBytes. The ceiling raised by the measured growth (+364 Darling, +353 Lite), with a change-log comment.
• Merged origin/dev. Resolved one conflict: both branches had raised TotalCeilingBytes from different base points, so the resolution kept both change-log comments and combined the growth.

Test plan

• Darling.Tests and Lite.Tests both build, 0 warnings, 0 errors.
• New live test DarlingMcpAnalysisFindingsBudgetLiveTests (30 planted chains, realistic prose, mixed remediation shapes). Default call is 28,711 bytes (budget 32,768), 18 of 30 chains returned. full_text: true, limit: 30 returns all 30 untruncated. The production "before" measurement (66,838 bytes) is documented in the test's own doc comment.
• Proved the pin is meaningful. Temporarily flipped the Darling tool's defaults back to old behavior (limit = 1000, full_text = true, parameters kept), rebuilt, and ran the live test. It failed (finding_count came back 30, not the pinned 18), then was reverted.
• Lite's twin live-shaped test in McpAnalysisFindingsCommandTests (same 30-chain seed) passes under budget.
• McpPayloadContractCensusTests: all pass (77 test cases in the combined Darling run touching this area).
• McpToolsListBudgetTests: both products, 5/5 pass, totals match pins exactly (Darling 172,083/172,083, Lite 90,234/90,234).
• DarlingWebEndpointsTests: 68/68 pass, including the new no-rig pin.
• Full Darling.Tests suite (rig-tg, port 55982, UTC): 13,882 total, 3 failed, 47 skipped, 1 not run. The 3 failures (TrendPayloadBudgetLiveTests, CaptureDownChunkOrderTests, ServerListAndSummaryPlanShapeTests) are TimescaleDB chunk-count/plan-shape tests unrelated to any file this PR touches (collection_log chunk append plans). They reproduce in isolation on this rig, but dev's own latest Build run (2026-09-19) is green. This reads as accumulated chunk state on the long-lived, reused rig-tg (shared across TG's session and mine), not a code regression. Flagging for the coordinator rather than filing an issue. I cannot rule out rig state without restarting a rig my brief told me to reuse as-is.
• Full Lite.Tests suite: 5343 total, 0 failed, 0 skipped. One run earlier in the session showed a one-off fail in AnalysisPassTokenThreadingTests.TheReadLockWaitIsAbandonableWhileAWriterHoldsIt, a timing-sensitive threading test in a file this PR does not touch. It passed both alone and in this final clean run.

CHANGELOG entry

None. This PR is test, description and wiring fixes on top of TG's still-unmerged #4198 work. The user-visible entry belongs with the parent #4198 PR once every lane's work lands.

Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3