fix(plugins/feishu): render markdown tables as CardKit v2 interactive cards

[chapaofan]

Summary

Fixes #50602

Feishu's post-type md elements do not support table syntax, so table-containing messages were force-sent as plain text, losing all table structure. This PR converts markdown tables to Feishu CardKit v2 table components.

Changes

1. _parse_markdown_table — Parse markdown text into segments of non-table text and structured table data (header + rows).
2. _build_table_card — Convert table data into a CardKit v2 table component with proper column definitions, row data, and header style.
3. _build_interactive_card_with_tables — Assemble a full schema 2.0 interactive card containing markdown text and table elements.
4. _build_outbound_payload — Send as interactive card when tables are detected, falling back to post/text as before.

Design decisions

• Cell content is stripped of markdown bold markers (**, __) since Feishu table data_type: text does not support inline markdown
• Table header uses header_style: {bold: true} for visual emphasis
• Non-table text segments between tables are rendered as markdown elements in the same card
• If no tables are detected, the original text/post fallback path is unchanged

Testing

Verified working with Hermes v0.16.0+ plugin architecture. Tested locally with real Feishu messages containing markdown tables — renders as native CardKit table component with proper column alignment and header styling.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #12114 — same Feishu CardKit v2 native-table rendering mechanism (helper names differ but the approach is identical), entering the saturated Feishu-table PR cluster (#17006, #48807, #48334 all point to #12114). Fixes #50602.

[teknium1]

Thanks for targeting a real Feishu formatting gap: current main still routes detected tables to plain text at plugins/platforms/feishu/adapter.py:4524-4530.

Problems

• _parse_markdown_table() classifies every outer-pipe line as a table (plugins/platforms/feishu/adapter.py:218) and _flush_table() emits it even when no GFM divider was present (:214). Literal pipe-delimited text will therefore become an interactive card.
• The parser has no fenced-code state (:217-245), so a code sample containing table-shaped lines is converted instead of preserved.
• _build_outbound_payload() is shared by send() and edit_message() on current main (:1904, :1958), while the PR always selects interactive for a recognized table (:4548-4550). The change needs an explicit streaming-preview/final-delivery contract and coverage for both paths.
• The PR modifies no tests despite adding parsing and delivery behavior.

Suggested changes

• Require a validated header-plus-divider sequence, preserve fenced-code content, and add parser plus send/edit regression tests.
• Define and test how streamed previews transition to the final CardKit payload before routing edits through the new path.

Automated hermes-sweeper review.

[teknium1]

This emits a table even when table_lines contains only | literal | or otherwise lacks a GFM divider. Require a validated header-plus-divider sequence before appending a table segment; otherwise return these lines to the text segment.

[teknium1]

Table detection runs inside fenced code because this parser tracks no fence state. Preserve pipe-shaped examples inside ``` fences as text/code rather than turning them into CardKit tables.

[teknium1]

This helper is also used by edit_message() on current main. Please make the streaming-preview versus final-card behavior explicit and test both call paths before returning an interactive payload for every detected table.