[liyuankui]

Problem

When annotating long blocks in Remark Mode, the line reference spans the entire block — e.g., selecting one cell in a 10-row table produces L28–L37. This makes the exported feedback hard for AI agents (and humans) to locate precisely.

Solution

Narrows the annotation line reference to the specific sub-element selected:

Block type	Granularity	Accuracy
Table rows	Per-row (exact)	Header = blockStart, tbody[i] = blockStart+2+i
List items	Per-item (approx)	Document-order <li> index within block (off by ≤1 for multi-line items)
Code blocks	Per-line (approx)	Newline count via text-node walk + 1 for fence line

Visual feedback: selected table rows and list items also receive a colored highlight (4 color variants each).

Architecture

Extracted findTrLineInBlock, findLiLineInBlock, findCodeLineInBlock, and narrowLineInBlock from the initRemarkMode closure into remark-utils.ts:

• Pure / testable — no DOM globals, duck-typed node stubs
• Single dispatch point — findBlockRange delegates through narrowLineInBlock for both single-block and cross-block selections

Tests

18 new unit tests added (55 total, all passing via node --test):

• findTrLineInBlock: thead, tbody row 0/1, non-table node, text-node input
• findLiLineInBlock: first/third item, not-in-list, nested list document order
• findCodeLineInBlock: no-pre block, offset 0, after 1 newline, second text-node, not-found
• narrowLineInBlock: table dispatch, list dispatch, plain paragraph (null), single-row

Design rationale

Precision is intentionally approximate for lists and code — the primary consumer is AI feedback loops where ±1 line is acceptable. The goal is to avoid exporting L1–L80 for an 80-line code block when the user selected line 42.