Make native query answers compact and honor output budgets

[forhappy]

Native query answers currently spend space on audit envelopes, duplicate node inventories and empty sections. Default CLI and MCP text now shows a compact, source-located answer with uncertainty intact; --verbose/--evidence exposes the full paged audit ledger. Raw query, Agent View and MCP structured contracts are unchanged, and exploration retains isolated requested symbols.

Finite commands share a strict UTF-8 output budget and an exact continuation. Commands without a native cursor save an immutable completed answer for compass output, so continuation never repeats an action. Oversized JSON/JSONL fails intact, and streaming events/watch/serve/REPL reject finite budgets before starting. Complete native responses use a bounded, corruption-checked cache keyed by verified graph, Program IR, profiles, request limits and semantic mode. Partial responses and source excerpts remain native.

Validation: workspace Clippy and native lib/bin tests passed (1,124 tests; two existing vendored grammar tests ignored), along with CLI product, all CLI suites (311 tests), renderer, MCP, store/query integration, CompassQL TCK/support and product-boundary gates. New regressions cover one-hop cost, Unicode budgets, immutable saved output, corruption/retention/redirected paths, machine output, init publication, cache invalidation and isolated exploration.

The fixed 50-question replay preserves all row verdicts (49/50) and reduces mean estimated output from 419 to 364 tokens, versus 550 for the installed comparator. The 25-question natural suite preserves 24/25 at 547 tokens versus 757. The qualification record includes source/artifact pins and both optimized binary hashes. On the four complete graphs, repeated-query medians are 30–34 ms versus 85–994 ms before; every raw JSON answer is byte-identical across binaries and repetitions. Partial Gson remains uncached, and some cold timings regress. These are source-reviewed text-recall panels over five pinned public repositories, not population-wide precision claims.

Compatibility and migration notes document compact defaults, the new output reader and --budget no longer selecting legacy traversal. JS/viewer and graph-publication qualification were not rerun locally because their contracts and source are unchanged; CI runs those broader gates.

All 14 CI jobs pass on 4cafb0eb, including the six native platforms, full CLI integration tests, nextest, query relevance, code-graph fixtures, JavaScript/VS Code, publishable archives, cargo installation and standalone launch, Windows installers, and dependency checks. Legacy fixtures explicitly select --traverse; disposable caches are excluded from Git and crate archives at every working-directory depth. Workspace packaging also passes locally with the four test-generated cache records present.