read_markdown improvements

What changed

Three related improvements landed together:

1. Adaptive output tiers

read_markdown now selects output detail based on file size:

For large files the response includes a must_follow instruction directing the agent to use the buffer ref for subsequent reads instead of the original path. This prevents repeated full-file reads for every heading navigation call.

2. @file_* buffer ref support

read_markdown now accepts @file_* buffer refs as the path argument for all subsequent calls (heading navigation, line-range slicing). Once a large file is loaded into a buffer ref, every subsequent access hits the in-process cache — no disk re-read.

{ "path": "@file_abc123", "heading": "## Configuration" }
{ "path": "@file_abc123", "start_line": 45, "end_line": 90 }

3. Heading navigation

Pass heading to jump directly to a section without reading the full file:

{ "path": "docs/guide.md", "heading": "## Installation" }

The response includes the matched section’s content, sibling headings (for orientation), and breadcrumb path.

Why this matters

Large markdown files (tracker docs, changelogs, architecture docs) previously required reading the full file to find a section. With adaptive tiers + heading nav, agents can navigate to the relevant section in one call and cache the buffer ref for the rest of the session.

Known limits

• @file_* refs expire when the MCP session ends or the output buffer is evicted (LRU, capacity 50). Re-read the original path to get a fresh ref.
• Heading match is case-insensitive prefix search — if two headings share a prefix, the first match wins.