fix: preserve Markdown context during JSON command recovery

[longchen-nv]

When JSON string recovery scans a string in isolation, it can reinterpret literal shell backticks from a surrounding Markdown code fence as inline-code delimiters. The example reported on #516 then changes from incomplete/CAUTION to complete/SAFE merely by adding unrelated array entries, allowing --fail-on-incomplete and MCP installation checks to accept unresolved static analysis.

Reuse the corresponding slice of the whole-document Markdown projection during recovery. This preserves the original delimiter ownership, source coordinates, JSON escapes, and disjoint-span work bound without reparsing each string as a new Markdown document.

Follow-up to the P1 review, as requested by the maintainer.

Regression coverage

• Add 30 cases covering single-entry/multiple-entry arrays, backtick/tilde fences, plain/blockquote/list/tab-indented list containers, and benign controls.
• Twelve public cases each invoke CLI and MCP, with semantic analysis disabled or successful deterministic model responses. Unresolved commands must retain static_parse_limit, CAUTION, strict CLI exit 1, and safe_to_install=false; benign controls remain complete/SAFE.
• Correct the existing fenced hostname example's unit and public expectations: a JSON code fence does not establish Markdown inline ownership inside its strings, so runtime-selected literal backticks remain incomplete even when surrounded by prose. Two additional public cases verify that the standalone inline-documentation control remains complete.

Validation

• Tests-first checkpoint: the initial 24 new cases produced 10 failures before the fix, including both public scan modes.
• Final focused run of all three affected test modules: 139 passed.
• Python 3.12 with locked development dependencies, make test-ci: 5,054 passed, 14 skipped, 38 deselected, four expected failures, 90% coverage. Its two failures were the existing fenced-hostname public expectations described above; those expectations were corrected, a standalone control was added, and the affected modules then passed the 139-case rerun. Production code is unchanged since the full run. Hosted CI will rerun the full suite on the final commit.
• Original review fixture through the actual CLI with --no-llm --fail-on-incomplete: score 0, CAUTION, is_complete=false, static_parse_limit, exit 1.
• Ruff lint, repository formatting, and git diff --check passed.
• Semantic-enabled tests use deterministic transports; no live LLM validation was performed. CodeRabbit CLI was signed out, so review used local code inspection and tests.

[rng1995]

[SkillSpector Review]

Reviewed current head b179f91ab7e9aa090a5fb4415a818f1c3c2cf9af, including the complete production diff, the whole-document Markdown projection and JSON-span recovery paths, all added unit/public CLI/MCP regressions, prior discussion, and exact-head checks.

Reusing the length-preserving whole-document projection correctly keeps each recovered JSON string's original fenced, literal, or inline-code ownership while preserving the validated structural-quote offsets and bounded disjoint-span work. The tests cover fence/container combinations, single and multiple entries, conservative runtime-selected cases, benign controls, semantic/static modes, CLI exit behavior, and MCP installation decisions. All five hosted checks pass. No required changes remain on this head.