fix: avoid false analysis limits in documentation

[mohgupta-ship-it]

Valid Markdown code spans and JSON request placeholders could produce false incomplete-analysis/AE1 results. Delimiters could also pair across unrelated Markdown blocks or table cells and hide unresolved runtime commands. This change establishes delimiter ownership from complete document context, preserving code bodies and original source coordinates.

• Recognize all strings in bounded, validated JSON values: object keys/values, array elements and scalar strings, including explicit JSON fences with attributes, list/blockquote containers and standalone JSON after frontmatter. Container indentation follows visual tab columns without expanding payload bytes. Invalid, partial or oversized containers grant no ownership; unknown fragment context stays conservative.
• Keep inline delimiter pairing within Markdown blocks and proven GFM table cells. Preserve ordinary multiline paragraphs and list continuations, distinguish table rows/headers from Setext headings and real block interruptions, and keep empty list items from creating phantom paragraphs. Retain fenced, indented and raw-HTML code handling, inline Make expressions and repeated documentation references.
• Recover runtime commands in disjoint validated JSON strings when an earlier parse would otherwise skip them. Retain the whole-content result, raw escapes and source evidence without reparsing overlapping suffixes. Forward JSON traversal and cancellation handling from fix: make JSON quote candidate scanning linear #521 remain intact.
• Preserve genuine removal-instruction findings and incomplete coverage for unresolved runtime commands. Proven JSON delimiters do not exempt their contents from analysis.

Completeness and risk remain separate: existing findings and thresholds are unchanged. Benign documentation can complete, while unresolved static inspection remains incomplete and fails strict CLI/MCP installation gates even when semantic analysis succeeds.

Validation on frozen source ca0ceb77e252c0ad5dc84eb435323edb80d28c57:

• Added 295 permanent tests across five modules, including exact reported inputs, paired benign controls, malformed containers, raw quote/source locations, table/JSON composition, parser-state transitions and counted-work/cancellation checks. Tests-first red checkpoints preceded the corresponding corrections. The focused suite passes 722 tests (295 new + 427 existing); an independent clean-source replay passes all 295 new tests.
• Local make test-ci: 4,793 passed, 14 skipped, 38 provider/integration tests deselected and four expected failures; 89% coverage.
• 80 permanent public test cases execute 160 CLI/MCP invocations with successful deterministic semantic transports. Enabled mode checks all three semantic analyzers and completed work; disabled mode checks zero calls. Incomplete inputs are rejected, while benign complete controls retain their expected installation decisions.
• Actual codex_cli replay: 40 fixtures in both modes, 80 CLI scans passing their expected completeness, evidence and exact strict/risk-threshold exit contracts. Enabled scans recorded 150 successful analyzer-level attempts out of 150, with all three semantic analyzers completed and no degraded scans. Disabled scans recorded zero attempts.
• Supplemental actual-provider MCP check: four exact reported fixtures in both modes, 8/8 passing, with 14/14 analyzer-level attempts successful and zero disabled-mode attempts. This exercises mcp_server.run_scan directly; it does not certify HTTP/stdio wire transport.
• Fresh Docker build and the unmodified local-skill/GitHub-URL smoke tests pass with --no-llm. All 97 installed Python source hashes match the frozen revision. Live-provider replay ran separately on the host using that same verified source.
• Repository lint/format, diff checks, secret scan, DCO and independent implementation review pass. Three pre-existing pytest timeout annotations were unregistered in the local environment; the new work/cancellation tests assert counters and callbacks directly.

The targeted and independent runs overlap the full suite; their counts are not additional unique tests. Provider telemetry counts analyzer-level attempts, not individual underlying model requests. These results supersede the earlier validation summary for b5c5d8e. A separate pre-existing legacy-backtick interpretation gap is not claimed fixed.

All five hosted CI checks pass on published ca0ceb77e252c0ad5dc84eb435323edb80d28c57: changes, lint, test-unit, DCO and docker-smoke (CI run). The Linux full suite independently matches the local result: 4,793 passed, 14 skipped, 38 deselected, four expected failures and 89% coverage.

Fixes #515.

Prepared by Codex on behalf of Mohit Gupta.

[rng1995]

[SkillSpector Review]

Reviewed current head bce6afadb2819d2235510e183eaa3a197ac02463 — APPROVE.

I reviewed the post-merge source diff, including bounded JSON ownership, Markdown block/table delimiter ownership, source-offset restoration, runtime checks, and the public completeness gates. The implementation remains conservative when context is incomplete, keeps command bodies analyzable, bounds validation and projection work, and has extensive success/failure/boundary coverage. I found no required changes.

The head changed during review from ca0ceb77e252c0ad5dc84eb435323edb80d28c57 to this merge commit, so the earlier assessment was discarded and the current diff was re-reviewed. Merge gate: GitHub reports mergeStateStatus=BLOCKED and no checks yet on the new head; required lint, unit, and DCO checks must pass before merging.

[mohgupta-ship-it]

Thanks everyone for code reviews. As I have got two approvals on this change. I wll merge it post the pipeline passes.