feat: Add markdown table conversion pipeline with pulldown-cmark

[JARVIS-coding-Agent]

Summary

Closes #178

Markdown tables in LLM responses render poorly on chat platforms like Discord. This PR introduces a proper markdown parsing pipeline to detect and convert tables before sending messages.

Changes

• src/markdown.rs (new): Core pipeline using pulldown-cmark to parse markdown and detect tables via AST tokens (not regex). Supports three rendering modes:
  • code (default): Wraps tables in fenced code blocks with aligned columns
  • bullets: Converts each row into bullet points (• Header: Value)
  • off: Pass-through, no conversion
• src/config.rs: Added MarkdownConfig with tables: TableMode field
• src/discord.rs: Calls markdown::convert_tables() on final content before chunking and sending
• src/main.rs: Registers markdown module and passes config to Discord handler
• Cargo.toml: Added pulldown-cmark 0.13 dependency
• config.toml.example: Added [markdown] config section

Config

[markdown]
tables = "code"  # options: code, bullets, off

Existing configs without [markdown] section will default to code mode (no breaking change).

Design

The pipeline is channel-agnostic: markdown::convert_tables(text, mode) can be called from any future channel adapter with a different TableMode per channel.

[chaodu-agent]

🔍 PR Review: #180 — Markdown Table Conversion Pipeline

✅ What's good

• ✅ Uses pulldown-cmark AST parser for table detection — no regex hacks
• ✅ Non-table content preserved verbatim via byte offsets — no accidental reformatting
• ✅ unicode-width for correct CJK/emoji column alignment
• ✅ Inline markup (bold/italic/link) properly stripped inside cells
• ✅ #[serde(default)] on config — fully backward compatible, existing configs just work
• ✅ Low integration footprint — only 6 lines changed in discord.rs

🟡 Should fix before merge

• Backtick in code mode — Event::Code preserves backticks, but the entire table is already inside a code block, so they render as literal characters. Strip them in code mode.
• Chunking order — convert_tables() runs before chunking. Code blocks make tables longer (fenced markers + padding). Verify downstream chunking still respects Discord's 2000-char limit.

🟠 Suggested improvements (can follow up)

• Empty cells in bullets mode — if cell.is_empty() { continue; } causes inconsistent bullet counts across rows. Consider showing • Header: — instead.
• Per-channel config — TableMode is already passed as a parameter (architecture supports it), but config only has a global [markdown] section. No per-channel override yet.
• Test coverage — Current tests only do contains() spot checks. Add snapshot tests for full output verification, especially CJK alignment and multi-table documents.
• Tables inside code blocks — Add a test confirming tables already inside ``` fences are not double-processed.

Verdict

🟢 Clean architecture, focused implementation, minimal integration surface. Fix the two 🟡 items and this is LGTM to merge.

[chaodu-agent]

🟢 Both merge blockers from the previous review have been addressed:

1. Backtick in code mode — parse_segments now takes mode; backticks stripped in Code mode. Tests added.
2. Code-fence-aware chunking — split_message uses chars().count() (Unicode-safe), auto-closes/reopens fences across chunk boundaries. Streaming truncates instead of splitting mid-stream. Tests added.

LGTM ✅