Affected Chromium adapters fail after selecting WebGPU even when packed renderer paths are forced. Use WebGL2 for the web app and let device limits select renderer tiers.\n\nRefs #788
dms() rounded the seconds after splitting the degrees and minutes off, so
seconds that rounded up had nowhere to go:
45.9999 deg at AUPREC 0 -> 45d59'60"
89.99999 deg -> 89d59'60"
Round once, into the smallest unit shown, and split degrees and minutes out
of the rounded total, as format_length now does.
surveyor() then needed the same treatment one level up: it chose its pole
and quadrant from the unrounded direction, so a bearing that rounded to the
quarter turn came out as the contradictory `S 90d0'0" E` the function's own
doc comment sets out to avoid. Snap the direction to the shown resolution
before deciding, so 359.99999 deg at AUPREC 0 is `E` rather than a bearing
of ninety degrees.
format_length rounded the fraction after splitting the whole part off, so a
value that rounded up to a full unit had nowhere to go:
fractional 5.995 -> 5 (64/64 reduced to 1/1, then dropped)
fractional 0.995 -> 0
architectural 11.999 -> 0'-11"
architectural 23.999 -> 1'-11"
engineering 11.99 -> 0'-12.0" (twelve inches is a foot)
engineering 23.99 -> 1'-12.0"
Round once, into the smallest unit that will be shown, then split feet and
whole inches out of the rounded total. The carry reaches every column, so
the same values now read 6, 1, 1'-0", 2'-0", 1'-0.0" and 2'-0.0".
Dimension definition points come out of a DWG as ordinary POINT entities on
the Defpoints layer, so the header's PDMODE applied to them like any other
point. Opening a drawing authored elsewhere with a point style set stamped
that glyph on every dimension end, burying the drawing under crosses and
circles.
They are a dimension's bookkeeping, not draughting the user placed, so
resolve PDMODE to 0 for them and leave the single pixel-sized dot. Both
render paths — the header one and the viewport-aware relative-PDSIZE one —
go through the same resolution, so a relative PDSIZE cannot bring the glyph
back.
Fixes#790
The MText editor paints its own caret and is not a focusable widget, so
the command input keeps widget focus behind the dialog. It captures the
arrow itself, which routes the key through CommandLineArrowProbe instead
of ArrowKeyPressed — and that path only asked whether the command input
was focused, never whether an editor was open. Up and Down therefore
walked command history while the user was typing in MText.
Give the probe the same rule ArrowKeyPressed already applies: while the
editor is open the arrow moves its caret. The probe now carries the shift
state so a held Shift still extends the selection on this path.
Fixes#655
solid_boolean() collects the selected solids straight out of
scene.selected, a HashSet, and then uses handles[0] and handles[1] as the
two operands. Set iteration order is arbitrary and can differ between
runs, so SUBTRACT — the one non-commutative operation of the three — could
cut either solid out of the other for the same selection.
Sort the pair by handle, so the solid created first is the body the other
is taken out of. UNION and INTERSECT are unaffected.
A circle of radius r tangent to a circle of radius r_i has its center at
r_i + r (external tangency) or |r_i - r| (internal). ttr_lc_candidates and
ttr_cc_candidates built that second distance as r_i - r and skipped the
combination when it came out negative, so every solution where the new
circle is the larger one and encloses an existing circle was dropped. TTR
over two circles could only ever return externally tangent results.
Only rho^2 reaches the algebra below, so taking the magnitude is enough,
which is how the tan-tan-tan path already computes it.
Fixes#814
The embedded-expression rescan in eval_to_string tries every substring, so
any identifier the evaluator knows is also recognised in the middle of
ordinary text. Because `pi` needs no arguments it resolves anywhere it
appears, and names typed into property or command fields were rewritten:
`ASPIRATORE` came back as `AS3.141592653589793RATORE`.
Require a candidate whose first or last token is an identifier to start,
respectively end, on a word boundary. Leading and trailing numbers stay
unrestricted so mixed input such as `def10*10abc` still evaluates.
Fixes#809
Saving to a path that already holds a drawing fails on Windows with
"could not verify <path> before saving: ... (os error 33)". Save As over any
existing drawing is affected; saving to a new path is not.
EditLease::acquire() opens the destination and takes the platform drawing
lock on it, then the save fingerprints the same path to check nothing changed
underneath. LockFileEx is mandatory and scoped to the handle that took it, so
that second open is refused against our own lock, lands in the non-NotFound
arm and aborts the save. flock on Unix is advisory and the second open
succeeds, so neither CI nor a Linux desktop sees it.
Read through the handle the lease already holds instead. A duplicate shares
the lock ownership, so the pre-replace re-check on the save worker can use
one too rather than re-opening the path.
The same collision reached the headless open op, which read a path this
session may hold leased: automation's save_then_open_round_trips fails on
main for that reason, one line after the save it no longer reaches.
A unitless block has no source length to convert from, so falling back to
MEASUREMENT gave it an inches or millimetres identity it never had and
scaled the reference on insert. The properties panel reported unit_factor
1.0 for the same insert.
Skip the conversion when either the block or the drawing is unitless.
Fixes#706.
Each ribbon panel was rendered five ways every frame so CollapsePanels could
roll back to any density at layout time: full, compact, button, tight and a
full-size flyout that duplicated the full rendering.
The layout pass ran four full measure-layouts per panel even in forced modes,
and the degradation decision was an untested u8 inline in Ribbon::view.
Now the choice is a pure, unit-tested decide_levels() whose levels are a typed
Level enum (Full/Compact/Collapsed/Tight). Panel carries its four renderings
as elements: [Element; 4] indexed by Level, with a slot(i, level) helper for
the tree stride. Forced modes (Full/Compact/Collapsed) measure nothing at all
and derive the row height from the placed panel. The open flyout draws the
panel's own `full` rendering inside a self-drawn box (identical weakest-bg +
1px neutral-border to the container it replaced), so no duplicate
render_group is built per panel per frame. Level, Panel, Widths, CollapsePanels
and decide_levels are pub(crate); CollapseMode stays pub (re-exported).
Tests (collapse.rs tests module, added RED first):
- forced_modes_pin_every_panel
- auto_huge_width_keeps_everything_full
- auto_degrades_from_the_right_one_panel_at_a_time
- auto_stops_degrading_once_the_row_fits
- auto_escalates_a_single_panel_past_collapsed_to_tight
- auto_cascade_drops_every_collapsed_panel_to_tight
- auto_runs_the_compact_phase_across_the_row_first
- no_panels_stays_empty
- single_panel_walks_the_whole_ladder
- zero_widths_never_overflow
- endless_row_degrades_everything_to_tight
- unbounded_row_keeps_everything_full
- nan_budget_degrades_both_phases_but_no_tight_cascade
- exact_fit_at_full_stays_full
- matches_the_pre_refactor_algorithm: byte-equivalence sweep against the
verbatim pre-refactor u8 algorithm (280 width/budget combos incl. the
degenerate 0/0/inf/NaN inputs), proving the decision logic is unchanged.
Measured with the committed #[ignore]d benchmark
(`cargo test --lib --release -- --ignored --nocapture bench_ribbon_view_construction`)
run as one instrument across pre-change / this change: Ribbon::view()
construction ~676 -> ~425 us/frame (~37% of view construction saved); forced
modes run 1 layout call per panel per frame instead of 5.
Verification: cargo build clean; cargo test --lib = 382 passed (367 baseline
+ 15 new ribbon tests) / 2 known pre-existing failures unchanged
(save_then_open_round_trips, tilted_ucs_places_planar_entities_with_the_plane_normal).
Hakan Seven