fix(analyzer): preserve inline JSON and nested list ownership

[zozozeezee]

Complete inline JSON values could leave their closing quotes available to the shell parser, causing ordinary documentation later in the file to exhaust its bounded word parser. Nested-list continuation text could also be treated as indented shell code, letting inline delimiters borrow operands from later blocks.

Reuse validated JSON-string spans within proven paired inline code, and account for a single list container when determining code indentation. Preserve conservative continuation handling for multiple list markers and unsupported table containers. Genuine shell constructs retain their existing limits and findings.

Validation: 246 Markdown table/prose tests pass; synthetic cases cover both false limits, destructive commands, nested fences, indented code, and oversized shell input. Full upstream suite: 13,684 passed, 14 skipped, 245 deselected, 4 expected failures. No release or deployment is included. Independent review found no remaining issue after correcting a nested-table regression caught by the full suite. Ruff and changed-diff secret scan pass. The same production patch passes the 8,021-test 2.12 backport suite.

[yashrajp22]

The inline JSON and nested-prose cases are fixed, but two Markdown boundary regressions remain: fences can make valid guides fail strict scanning, and nested tables can lose required incomplete-analysis results. Both reproduced in source and newly built, noneditable wheel installs.

Validation: 380 base and 387 HEAD tests passed in each mode; all 96 offline CLI scans completed with source/wheel agreement. Existing partial corpus results remain explicit. This is targeted verification, not a claim of global accuracy or live-provider coverage.

[yashrajp22]

Could we update fence ownership when a list ends? A top-level fence immediately after a list inherits indentation 2 here, so its column-zero closer is ignored. For example, append 500 lines of Ordinary documentation. to this prefix:

- Item
```sh
printf ok
```
The interface doesn't match the spec.

The referenced guide completes with exit 0 on base, but HEAD reports static_parse_limit/AE1 and exits 1 with --fail-on-incomplete, in both source and fresh-wheel scans.

Source inspection also shows the related reverse transition: a valid nested fence such as ~~~sh stays active when a nonblank line leaves its list. Please clear stale list ownership before opening an outdented fence, and end an existing list-owned fence when its container ends. The latter variant was checked against source, not separately executed.

[yashrajp22]

Could we preserve table-cell boundaries when allowing this deeper list indentation? With list markers on separate lines, code_indent becomes 8, but _markdown_table_delimiter_columns still rejects a delimiter row at column 4. This valid nested table therefore enters ordinary paragraph handling and pairs backticks across different cells:

- Modes:
  - | Left | Right |
    | --- | --- |
    | `$(resolve_tool).example | ` -rf / |

The same issue occurs across table rows. On base these referenced guides retain static_parse_limit/AE1 and fail strict scanning; HEAD reports complete analysis and exits 0 in both source and fresh-wheel scans. Please make table recognition use the same container indentation, or retain literal delimiters when that ownership is unproven. The existing nested-table test uses - - on one line and misses this shape.

[mohgupta-ship-it]

i see but it depends on the identation one follows. can cause FPs too?

[mohgupta-ship-it]

Powered by Codex: reviewed on Mohit's behalf.

NEEDS_CHANGES — P1 and P2 independently reproduced. Rating: bugfix; Tier 1. Frozen head edc2cc1dc1b5ca60f79cf84c5fdede862101ddbb, base 240421219409dc0bafef9f618a10d80c2d7e83a3.

1. P1 / high confidence — nested tables lose incomplete-analysis evidence. The changed list-relative cutoff at src/skillspector/nodes/analyzers/static_patterns_tool_misuse.py:4739–4741 permits deeper paragraph handling, while the table delimiter helper at 4626–4627 still rejects absolute indentation >=4. Separate-line nested lists containing valid tables consequently pair backticks across separate cells/rows. Both strict CLI cases change from base AE1 / incomplete / exit 1 to HEAD SAFE / complete / exit 0. Independent Markdown parsing confirms actual table cells, so the indentation concern raised in the existing thread does not invalidate these cases. This corroborates the existing table finding; it proves lost completeness signaling, not harmful execution.

2. P2 / high confidence — stale list ownership makes valid guides fail strict scanning. Lines 4852–4859 retain list indentation when opening a top-level fence; closure at 4809–4821 also fails to end list-owned fence context on container exit. Both the list-to-top-level-fence case and the reverse nested-fence-to-outdented-prose case change from base success to HEAD AE1 / exit 1. The reverse case was executed independently as well, corroborating the existing fence finding.

Repair table recognition relative to the actual container, never pair delimiters across cells/rows, and reset fence/list ownership at container transitions. Add projection, completeness-ledger and strict-CLI tests for both table variants and both fence transitions.

Validation: 1,061 existing test executions passed across source base/head; 12 paired offline CLI cases completed, with network connections blocked. The intended inline-JSON fix works and the concrete destructive-command control still emits TM1. I also reproduced the table failure on the exact wheel adopted by evaluator !4. That backport comparison introduces the table failure relative to 2.12.0; its fence failures were already present in 2.12.0.

Current head is unchanged, review remains changes-requested, and no hosted status checks are exposed. These are targeted offline results, not live-provider or global-accuracy certification.