D1185 — Tool results are lighter by default: compact JSON, search limit 8 and no related, no frontmatter_raw

Amended by D1354: the defaults of search go back to 20 hits and a related block of 5, and the budget rule is replaced by a measured average plus a margin; compact JSON and the conditional frontmatter_raw stay.

Decision. Every JSON tool result is built by jsonResult, compact (no indentation), and a test forbids MarshalIndent in internal/mcpserver. search returns 8 hits by default and a related block only for related: N. concept_read sends frontmatter_raw only when the frontmatter does not parse; the exact bytes are content under with_content.

Why. Measured on 20 real agent sessions, a search answer weighed 4.2 KB on average: the related block was 19% of it and the indentation 18%. The same questions asked with limit: 8 and related: 0 averaged 2.2 KB, 20 of 20 on a blind-judged answer set, with no loss of quality. In concept_read the frontmatter travelled twice (3.5%). The saving a client sees comes from fewer calls, not from less text per call, so the plan trims only what no answer needs and adds no nudge to read by section. The price is a breaking change for a client that relied on the old defaults or on frontmatter_raw in a default read.

Alternatives rejected. - Compacting in callTool whatever looks like JSON: some tools return stored text verbatim (index_get without with_hash, artifact and asset content), and a stored file that happens to be valid JSON must stay byte-identical; compaction is at construction. - A with_raw flag on concept_read: one more parameter in every agent's tools/list for a field nothing reads. - Dropping frontmatter_raw in every case: loses the repair path of a page whose frontmatter does not parse. - Clamping an explicit search limit: not needed, the Atlas endpoint has its own default and clamp.

Consequences. Amends D185 and D309 (frontmatter_raw is no longer unconditional), D781 (the default of related is 0; the parameter stays) and D70 (snippets were sized for limit 20). TestToolResultGolden proves a result equals its golden up to whitespace; TestSearchDefaultAnswerBudget keeps the default search answer under 2.5 KB on average, and if it fails the number is not raised; the generated guidance never tells an agent to pass limit or related to search or to read runbooks, procedures and incidents by section. The Atlas search endpoint keeps its own limit (20, clamped to 100) and related: 0.