feat: block-level and inline markdown rendering for LLM responses

[KUSH42]

PR: block-level markdown rendering

⚠️ Stacked on feat/rich-output-renderer (inline markdown PR) — merge that first.

Branch: feat/markdown-block-rendering

Summary

• Adds apply_block_line(line) — single-line-detectable block elements rendered to ANSI
• Extends apply_inline_markdown with images, links, and HTML inline tags
• Chains apply_block_line before apply_inline_markdown in both format_response and the streaming path

What's rendered

Element	Input	Output
h1	# Foo	bold bright-white, # stripped
h2	## Foo	bold white
h3	### Foo	bold
h4–h6	#### Foo	bold dim
Horizontal rule	--- / *** / ___	dim ─ line across terminal
Blockquote	> text	dim ▌ gutter + dim text
Nested blockquote	>> text	collapses to single ▌
Unordered list	- item	• at depth 0, ◦ at 1, ▸ at 2, · at 3
Ordered list	1. item	unchanged (already readable)
Reference link def	[ref]: https://…	suppressed (metadata, not prose)
Image		[img: alt] dim placeholder
Link	[text](url)	underlined text, URL discarded
<em>	<em>text</em>	italic, tags stripped
<strong>	<strong>text</strong>	bold, tags stripped

Architecture

apply_block_line is purely line-local — no lookahead or state. Two early-exit guards:

• "\x1b" in line — already ANSI-rendered, skip
• "\n" in line — multi-line block from StreamingBlockBuffer (PR4), skip

Headings and blockquotes apply inline spans internally (with reset_suffix set to the heading/gutter style) so a closing \033[0m from a span doesn't kill the outer colour.

List lines do not inject ANSI, so the outer apply_inline_markdown call handles their inline spans naturally.

format_response(text):
  pass 1: fenced code blocks → ANSI highlighted   (existing)
  pass 2: per non-ANSI line:
            apply_block_line(line) → apply_inline_markdown(result)

Streaming (_emit_stream_text / _flush_stream):
  out is line → apply_inline_markdown(apply_block_line(line), reset_suffix=_tc)

Files changed

• agent/rich_output.py — apply_block_line, extended apply_inline_markdown, updated format_response; removed dead _MD_OL_RE; fixed _MD_REF_LINK_RE to suppress titled reference-link definitions
• cli.py — import apply_block_line, chain in _emit_stream_text and _flush_stream
• tests/test_rich_output.py — TestApplyBlockLine (23 tests), 5 new TestApplyInlineMarkdown cases, 3 new TestFormatResponseInlineMarkdown cases

Bug fixes included

• format_response newline loss — splitlines(keepends=True) fed newline-bearing strings into apply_block_line whose capture groups stop at \n; matched block elements (heading, hr, blockquote, list) were returned without \n and concatenated with the following line. Fixed by switching to splitlines() + manual "\n".join with trailing-newline restoration.
• Blockquote reset_suffix — inline spans inside blockquotes were resetting to terminal default instead of restoring the dim gutter style. Fixed by passing reset_suffix=_BLOCKQUOTE_ANSI to apply_inline_markdown in the blockquote branch (mirrors the existing heading branch pattern).
• _MD_REF_LINK_RE titled forms — the $ anchor prevented suppression of reference-link definitions that include a title ([ref]: url "Title", [ref]: url (Title)). Removed $; the URL \S+ is sufficient to distinguish ref-links from prose.

Known interim regression

--- immediately after a paragraph renders as an hr instead of a setext h2. Corrected in PR4 (StreamingBlockBuffer intercepts the setext case before apply_block_line sees it; see docs/spec-markdown-stateful-blocks.md). Setext headings are rare in LLM output; the interim behaviour is acceptable.

Non-goals (deferred to PR4)

• Setext headings (=== / --- underline)
• Tables
• Multi-line blockquote continuation (no > prefix on continuation lines)

[KUSH42]

Superseded by NousResearch#4501