feat: make CLI and MCP output lean by default

[DeusData]

What does this PR do?

Makes CLI and MCP responses lean by default while preserving the graph answer itself.

• keeps core graph rows, exact totals, truncation state, and continuation cursors in the default response
• moves diagnostics, evidence, progress, and other operational detail behind explicit flags/parameters
• adds semantic-boundary output budgets and lossless, snapshot-bound pagination instead of byte-slicing answers
• adds response-local directory-prefix dictionaries only when the serialized and token-shape estimates show a material win
• keeps JSON as an explicit opt-in and adds a quiet CLI mode
• preserves invalid UTF-8/reversible identities instead of replacing bytes
• scopes code search to canonical indexed files and keeps scoped file globs portable to BusyBox/Alpine
• documents the lean/default versus detailed contracts and adds regression/size contracts

Measured output reduction

On the same serialized MCP tools/list fixture:

• bytes: 23,075 -> 15,071 (-34.7%)
• cl100k tokens: 5,068 -> 3,478 (-31.4%)
• o200k tokens: 5,151 -> 3,570 (-30.7%)

For repeated-path result tables, response-local directory compression reduced token-shaped output by 38.3% on Unix paths and 44.5% on Windows paths. Scattered/token-hostile inputs remain uncompressed, so the optimization does not make those responses larger.

Output contract

• defaults are compact, but not lossy: graph identities/rows, exact counts, safety signals, and continuations remain present
• max_output_tokens is enforced as a deterministic model-neutral byte ceiling at whole semantic units
• omissions are explicit and independently pageable
• cursors are bound to query arguments, result state, and index generation; stale cursors fail closed
• directory prefixes are response-local, non-recursive, and emitted only above exact savings thresholds

Validation

Final commit:

• env CBM_NO_CCACHE=1 scripts/test.sh --suites mcp: 253 passed, 2 expected skips, no sanitizer findings

Completed before the final two-file BusyBox portability follow-up:

• macOS full sanitizer suite: 7,537 passed, 4 expected skips
• Linux arm64 sanitizer suite: 7,375 passed, 4 expected skips
• container lint/static-format gate: passed
• deliberate RED-on-revert checks bound the CLI quiet/runtime, MCP output, search completeness, and watchdog-isolation regressions

The wider final rerun and native Windows completion were stopped to publish the PR; hosted CI is expected to validate the final merge tree. The prior Windows attempt failed in test infrastructure before product tests because a concurrent VM cleanup removed its live temp root; the affected infrastructure files are unchanged by this branch.

Checklist

• Every commit is signed off
• Focused tests pass locally on the final commit
• New behavior is covered by regression and output-size contracts
• Documentation is updated