feat: granular text selection + workflow on-response example - #55

…ode#55)

Two items from the deferred audit list.

## Granular text selection

Selection model was line-only: clicking a chat line selected the
whole line, dragging extended row-by-row. Users couldn't select a
substring within a single line (a paragraph fragment, half a
filename, the right half of a diff hunk).

Selection anchors are now `(buffer_line_index, char_offset_in_line)`
pairs instead of `Option<usize>`. New `Renderer::buffer_pos_at(row,
col) -> Option<(usize, usize)>` maps mouse coordinates to a buffer
position, accounting for `content_indent()` and clamping at
end-of-line so dragging past the right edge anchors there.

The viewport paint splits each row into up to three painted runs
(pre-selection, selected with `Attribute::Reverse`, post-selection)
so highlighting happens per-character within a line. Bold-glow
for bright colors is preserved across all three runs.

`selected_text` extracts:
- single-row: char slice [start_col, end_col)
- multi-row: tail of start row + full middle rows + head of end row

Reverse drag (end before start in row-major order) is normalized
so the extracted text is always the visible selection.

Mouse handlers (`MouseDown`, `MouseDrag`, `MouseUp`) now pass
`col` through to `buffer_pos_at` instead of dropping it.

UTF-8 safe: char counts use `chars().count()` everywhere, so
`é` and `🦀` each count as 1 char, not their byte widths.

## workflow.janet on-response example

New `plugins/response_inspector.janet` demonstrates the
`on-response` hook's two distinct capabilities:

1. **Pattern detection + notifications** — calls `harness/notify`
   when the agent's reply contains a code block.
2. **Steering string return value** — returns a system-prompt
   suffix to be applied on the next turn, asking the agent to
   add inline comments when its previous code block lacked them.

Documents the difference between `on-response` (assistant text)
and `on-tool-end` (tool output, where `harness/replace-result`
belongs).

## Test plan

- [x] 6 new tests in `ui::renderer::tests`:
  - single-row substring selection
  - reverse-drag normalization
  - multi-row spans (tail + middle + head)
  - empty selection returns None
  - UTF-8 char-index correctness (`café 🦀`)
  - `buffer_pos_at` clamps past-EOL
- [x] `cargo test --features plugin` -> 612 pass, 0 fail.
- [x] Both build profiles -> 0 warnings.
- [ ] Eyeball: drag in chat to select a substring; verify
      reverse-video highlight only covers the selected chars,
      clipboard receives the right substring.

Co-authored-by: Yogthos <yogthos@gmail.com>