[kirillDevPro]

Fix

Filed by the fix lane.

Fixes #127

TreeMap::remove_range(start, end) with end before start re-seeked the
cursor from the tree's first item (Cursor::seek resets), so the suffix from
end was appended after the slice up to start and every key in between was
duplicated. Per the owner decision on #127, a reversed range is empty: the
cursor now only seeks forward when end is at or past its position, otherwise
the map stays as it was, with no panic. Non-reversed ranges take the same path
as before (seek_forward from the slice end lands where a fresh seek did).
The only callers are the crate's own tests; nothing outside moon-sum-tree
calls remove_range.

Test: remove_range_with_end_before_start_leaves_the_map_unchanged in
crates/moon-sum-tree/tests/tree_map.rs failed before (left
[0, 1, 2, 1, 2, 3, 4], right [0, 1, 2, 3, 4]) and passes now.
Re-run: cargo test -p moon-sum-tree --test tree_map remove_range.

Reviewed by Codex: no concrete defects found.

Filed by the fix lane.

Fixes #129

When a GPU canvas's prepare_text() failed, the fallback in
Window::frame_gpu_canvases checked cached_env() instead of reusable_env(),
so a cached text frame holding polychrome (emoji/image) sprites was
replayed even though Window::drop_image may have freed those atlas tiles.
The fallback now applies the same rule as the reuse path; cached_env() is
left test-only since nothing else calls it. The issue's code path and
recommended fix were confirmed against current master as written.

Test: window::tests::err_never_replays_polychrome_text
(crates/moon-gpui/src/window/tests.rs) failed before (a polychrome sprite
reached the scene after the failed prepare: left 1, right 0) and passes now.
Re-run: cargo test -p moon-gpui --lib err_never_replays_polychrome_text

Reviewed by Codex.

Docs

What was stale: GPU_CANVAS_TEXT_API.md said every UI redraw prepares the
text of all canvases and that prepare_text() runs only for a frame that
will actually be shown. Since #146, Window::frame_gpu_canvases prepares no
canvas text at all while PlatformWindow::can_present() is false, even on a
UI redraw, and Window::present keeps needs_present armed when the backend
rejects the present (e.g. DXGI reports the window occluded).

What the code does now: while the window cannot present, no canvas runs
prepare_text(); the owed frame stays owed, and the first tick that can
present again passes that debt as force_present, so text_action prepares
every canvas's text. Invariants 2 and 3 and the upgrade checklist now say so.

How verified: read Window::frame_gpu_canvases (presentable && gpu_canvas_prepare_text(..)), the frame callback, Window::present
(needs_present.set(!accepted)), text_action in window/gpu_canvas_text.rs,
DirectXRenderer::present, and the regression test
deferred_present_skips_text_while_unpresentable_and_finishes_on_restore.

Reviewer: Codex. It confirmed the added sentences and pointed out that the
existing "a frame that will actually be shown" overstated the guarantee,
since the backend can still reject the present; that sentence now says so.

No issue opened.

Hunt

Input: a public builder that ignores its argument and says so in a
whole-line comment:

pub fn tone(self, _tone: MoonTone) -> Self {
    // `_tone` is accepted for API parity and ignored for now.
    self
}

Wrong output: function_body_uses_param returned true (the comment line
contains _tone), so scan_noop_markers did not count the builder and
noop_public_api_markers stayed flat; the audit passed on a builder that
leaves the component unchanged.

Expected output: false, i.e. the builder is counted. Source: the
scan's own contract ("builder methods that accept parameters but leave
the component unchanged") - a comment changes nothing - and the comment
rule the raw-hex scans in the same file already apply (a trimmed line
starting with // is skipped).

Test: function_body_usage_ignores_line_comments_naming_the_param in
xtask/src/component_audit/source_metrics/tests.rs failed before with
assertion failed: !function_body_uses_param(&unused, 0, "_tone") and
passes now. Re-run: cargo test -p xtask function_body_usage
cargo xtask component-audit --check-baseline still passes (no builder
newly counted).

Reviewer: Codex (PASS, no objections).

Leftovers: quiet (no open lane pull requests).
Trusted pull requests: quiet (none open).