The router began as a port of FreeRouting and has grown its own engine
(negotiated congestion, tree nets, exact checks, axis escapes), and
'freeroute' was one letter from the project it no longer follows. The
package, module and command are now rpmroute; the freeroute command stays
as an alias, and the RPMROUTE / RPMROUTE_ORACLE_JAVA variables fall back
to the old FREEROUTE names. Version 2026.10.07. Routing is unchanged: a
PCBench board routes to a byte-identical session either way.
With items away from the window left off its board, a stub from the window
edge out to a rerouted net's far pin crossed copper the exact check never
saw: 38 shorts on th-eda's flight controller. A connection with an end
outside the bounds is now left unrouted, and a window whose new copper
leaves the window is refused.
A window's board now holds every pin and piece of copper of the nets it
reroutes and, of the rest, only items near the window, instead of the whole
board rebuilt per window. The smallest windows go first, at most
--window-cap per round (40), and no window starts once --window-budget of
the routing's own time is spent (0.25). On th-eda's flight controller the
stage had taken 12x the routing time for no gain.
After the exact engine, every net routed wholly inside a window around a
stuck connection's pins is ripped, the rest of the board is laid down as
fixed copper, and the window alone is routed again at --window-resolution
(default 4) with the stuck net and the ripped nets only. The result is kept
only if fewer connections are left unrouted; each round doubles the margin.
The grid and exact routers gain only_nets and bounds for it. Off by default.
Anchoring a T only on the connection whose copper it touched missed the
connections that tie that copper to the T's own group. When one of those
was ripped, the T stayed routed but cut off: on charge-sense freeroute
counted 15 unrouted while KiCad found 27 opens. Routed connections of a net
form a forest, so the path back is unique; the T now depends on all of it.
KiCad's opens and freeroute's count agree again (18 against 17, the extra
one a GND pour island left to the fill).
A net's pin groups were joined as a chain in pin order, and each
connection had to end on one chosen pin. Groups are now paired by a
minimum spanning tree (Manhattan distance between their nearest pins), and
a connection may start or end on any copper its net has already routed on
that side (path cells, via cells on every layer; not walled cells, whose
via moves into the pad). Plane nets keep their drops to the plane.
A connection that tees onto another records it as an anchor. Ripping a
connection rips what hangs on it, transitively, inside the same rip-up
transaction, so rollback restores all of them. The exact engine gets the
connections in commit order with their anchors and drops a T whose anchor
it rejects. --no-tree restores the chain.
A QFN placed at 45 degrees had no exit on the grid: every horizontal or
vertical step from a pin passes its neighbours inside clearance. A pin whose
pad is turned more than ten degrees off the grid axes now leaves along the
pad's own axis, away from the part, until an L from there to a grid cell
clears every other net's pad, then searches as usual. The slanted leg is
checked and stored as an exact slab (segment_slab), not its bounding box,
in the grid stub checks and in the exact engine.
Where pins leave side by side, a cell names only its first owner, so stubs
are also indexed per cell for every connection near them; the exact
stub-vs-stub check, trace entry and rip-up all consult that index. A pin
with an axis exit and no clear cell reports no way out instead of falling
back to a cell on another net's route. diagnose.py replays with the exits.
Every connection is routed N times on the empty board, where routes may
share cells at a price that doubles each round for every other net on a
cell, and a cell that stays shared gains history. The real routing then
pays that history on free cells, so the first routes stop taking the space
later routes (a fine-pitch part's escapes) have no way round. Off by
default.
The old history term never steered: it only accrued at rip-up fights,
after the first routes had claimed the contested space, and routed
connections are never revisited in later passes.
Two connections of a net can share a via. The grid kept one owner per via
cell next to the claim list, and the search and rip-up read only that owner:
a route of another net through the via ripped the first user and left the
second's via under the new trace. The exact check then threw one of them
out. The search and rip-up now read the claim list, and the single owner is
gone.
The exact check also tested a via against the outline as its square box,
whose corner reaches past a 45-degree chamfer where the round via does not.
The grid keeps the via's centre its radius plus the edge clearance from the
outline, as KiCad measures it; the exact check now does the same.
kitspace_minisumo_v3 seed 4: 4 rejected connections to 0, 11 failed to 7.
A circle pad was its bounding box in both the grid and the exact check, so a
4 mm mounting hole's square corners refused legal routes 0.8 mm clear of the
real copper. It is now a 16-sided polygon whose sides touch the circle, so
the router is never closer than KiCad measures. diagnose.py takes --seed.
Rotated oval pads were built from padded corners that are not in outline
order, so they fell back to their bounding box; neighbouring boxes of a
rotated socket overlap, and the grid's own-pad exemption opened the
overlap to the pin's net, which the exact check then refused. Pads are now
the convex hull of their corners, the exact tree checks a rotated pad
against that outline, and the grid blocks cells against the same outline.
A via reused by a second connection of its net left that connection with
no claim on the via's footprint, so ripping the first owner freed cells
around a via still on the board. Footprint cells now keep every claimant.
A net class wider than the default was searched as a default trace: its
spread was empty below half a cell, halos were circular while the exact
check compares boxes, and a rip-up route only counted what its centre
crossed. The spread now covers the extra width in whole cells, halos are
square, and rip-up victims include what the spread crosses.
bench/diagnose.py now routes with net-class widths and hole spacing, as
the CLI does.
Copies the board (stripping its tracks unless --keep), exports the DSN
through mckicad, routes, applies the session and runs DRC with zones
refilled, counting hole problems whatever their severity. FREEROUTE and
MCK choose the builds under test.
A 1.0 mm trace from a 1.27 mm pitch pin (the RN2483 row on the TTN
Enschede mote) has to stay within 0.06 mm of the pin's centreline until
it is past the neighbouring pads. No grid column is that close, and the
router could not get there:
- A stub ran as one slanted segment to its cell centre. The exact check
sees that as its bounding box, which lands on the next pad. Orthogonal
stubs now leave straight along the pin's axis and jog to the cell
only at the end.
- End nodes were chosen as in free space even in the rip-up search, so
once another net crossed a pin's only exit, the pin had no end left
and rip-up could never take the exit back. In rip-up mode those cells
and stubs are now enterable at the rip penalty, and routes on a stub
are ripped along with the ones the path crosses.
- Two neighbouring pins' stubs each claim the cells a centred trace
would crowd, so their claims overlapped well before their copper
did. Stub against stub is now checked on the copper.
- Escapes also look one column either side, so two pins leaving the
same way can jog apart.
TTN Enschede mote, exact engine with --shove: 12 unrouted before, 0
after; KiCad DRC on the result has no clearance violations and no
unconnected items. The 0.5 mm pitch row test at resolution 1 now
routes all 5 pins (it pinned 2).
Committed vias were checked for hole spacing, but a route's own vias are
not committed until it is done, so a route hopping layers (more of them
on 6 layers) could put two of its vias 0.41 mm apart, as th-eda's
6-layer flight controller showed. A via is now also checked against
layer changes earlier on its own route.
A fixed via sits anywhere, not on a cell centre, and may have a bigger
drill than the router's via; checking hole spacing between cell centres
with one drill let the flight controller's routed 0.2 mm via come 0.19
mm from a hand-placed 0.25 mm one. The board now carries each fixed
via's drill (from KiCad's padstack name), and cells are checked against
fixed vias exactly.
A route could drop a via next to (or onto) a through-hole pad of its own
net, where it changes layer for free, and a via reusing a fixed via's
cell came out a little off the fixed via's centre, beside it. No via now
goes within hole spacing of a through-hole pad of any net; a via in a pad
is checked exactly against every via and hole; and a routed via on a
fixed via of its net is not written. charge-sense: 4 hole_to_hole to 0;
flight controller: 2 to 0, unconnected unchanged.
Fixed tracks and vias only kept other nets away, so a pin fanned out to
its plane was routed to the plane again, beside its own fan-out. Pins,
zones and planes that fixed copper of their net touches now form one
connected group: a fanned-out supply pin needs nothing more, a fanned-out
signal routes from its via, and copper touching no pin or zone adds
nothing to route.
The exact engine rebuilds its output from each connection's own vias,
so connections sharing a plane via each wrote a copy: th-eda's flight
controller had 71 spots carrying 2 to 4 vias of one net. Vias are now
written once per net and spot, and never on a via already on the board.
Every pin of a supply drops its own via to the plane, and nothing kept
two of them apart: the flight controller came back with 199 co-located
holes and 109 hole-to-hole errors, nearly all VDD_3V3 and GND. A via now
keeps its drill plus --hole-to-hole (default 0.25 mm, KiCad's) from
every via centre, fixed ones included, except an own-net via at the same
cell, which it reuses; such a via is written once, and none is written
on top of a via already on the board.
The grid is laid out for the default width; a wider trace also needs the
cells its extra half width reaches to be free, and claims a wider halo.
A connection that does not fit at its class width (typically at a
fine-pitch pin) is routed at the default instead and reported. A net is
written at its class width when all of its connections went in at it,
and the exact check verifies it at that width; nets that shove or the
diagonal pass reshaped go out at the default.
A start that was already a plane end joined the plane at once, with no
check that its via fits: under a sensor's via keepout the via went in
the pad and the exact check refused it. A via leaving or reaching a
walled-in pin goes in that pin's pad, but was only checked at the grid
cell's centre; it is now checked exactly at the pin against via
keepouts and other nets' pads and fixed copper.
Tracks and vias already on the board were ignored, so other nets could
be routed straight through them (the flight controller's 82 GND
stitching vias and its guard tracks). They are now copper of their net:
other nets keep a pad's distance from them in the grid, the exact check
and the room engine, and they add no connections to route.
A plane end is marked with its zone, not a pin, so a route from a pad
outside the plane's outline crashed looking for the pin's location
(the flight-controller board's first placement).
Over 8 net orders it left 25 to 30 fewer connections unrouted on the
dense sense boards and 16 to 18 fewer on two PCBench boards, and was
never clearly worse. KiCad DRC on the sense harness: 41 opens to 28.
The connections a rip-up route crosses are rerouted at once in free
space. If that leaves fewer connections routed than before, the new
route comes out and every displaced connection goes back on its old
path, with its rip count as it was. Even trades are kept.
A route may not leave a cell inside a wire keepout as a wire, nor reach
a pin there by one: it changes layer instead, and the via goes in the
pad. A pin whose every stub enters a keepout is reached by a via only.
A structure keepout named rule_area... (mckicad names KiCad rule areas
so) is checked for overlap only, as KiCad's DRC does, with the free ends
of each wire tested for their round caps.
--access-cost C makes a route pay C to pass where it would close another
pin's way out (its end cells, the via it may need there, and the
clearance around both) while that pin still has connections to route.
A pin inside a wire keepout can only leave by a via; its stub to the
cell centre was track inside the keepout, which the exact check refused.
The via now goes on the pin itself. Pin stubs are also checked exactly
against wire keepouts, as the exact engine will check them.
bench/diagnose.py reports what each refused connection collided with.
rejected by the exact check, unroutable even alone, walled in (a pin's
free region is closed off by other copper, with the nets that close it),
or congestion. The board state is the best pass's, rebuilt by replaying
its paths.
Each round routes every connection that has failed so far ahead of the
rest, which keep their order. The promoted set only grows, so two groups
cannot take turns failing; the best round is kept.
Runs every board with every variant under the same seeds and compares
them seed by seed, so a change is judged against the routing-order
noise rather than one lucky or unlucky run.
The router is deterministic, but its result depends heavily on which
connection routes first. A seed shuffles that order reproducibly, so a
change can be judged over several orders instead of the one it happened
to get.
On a four-layer board with a GND plane on an inner power layer, freeroute
ignored the plane (it only used zones on routing layers) and chained every
GND pad to the next with tracks across the board. Now a through-hole pad
inside a plane is joined to it already, and for a net with a plane every
other group connects straight to the plane: a route may end on any cell
inside it where a via fits (checked lazily, like other vias), and ending
there places that via. Signal nets are unaffected.
Tested on a four-layer DSN: three SMD GND pads each get a short stub and
their own via, the THT GND pad needs nothing, and no track crosses the
board. Ahead of the TerraHawk ESC, a four-layer board.
Recommend the exact engine for KiCad boards; document --edge-clearance,
--grid-resolution, --heuristic-weight and the unrouted report; list what is
read from a KiCad DSN (pads, vias, keepout kinds, outline, zones, fine
pitch), each checked against KiCad; replace the scope section with the
TerraHawk measurements and what still trails the JAR.
On the rc5 sense board at its real 0.2 mm clearance, ten pins of the
0.5 mm pitch STM32 could only reach a pocket of seven or eight cells
inside their own pads: the grid columns miss the narrow legal band past
the pad tip, so twelve connections could not route even on an empty
board. For a crowded pin (a neighbouring pad of another net closer than a
trace can pass) whose nearby cells are such a pocket against pads and
keepouts, offer straight stubs along each axis to the first open cell,
accepted only if the stub clears every other net's pad by the full
margin, measured exactly. All twelve then route alone; on the full board
unrouted connections drop from 61 to 59, the rest being congestion.
freeroute exited 0 and said nothing when it dropped connections, which
cost real diagnosis time: on rc4 two simple two-pin nets came back with no
copper and no explanation. The grid now records each connection it could
not route, the exact check adds each it rejects, and the CLI lists them on
stderr (the SES is still written; the exit code stays 0, as the JAR's).
A connection whose ends both sit in one zone of its net is listed apart,
as left to the zone fill: freeroute does not route through a zone it
cannot trust to be solid, but KiCad's fill often joins those pads anyway.
Resolution 2 reaches 0.5 mm pitch pins and cut unconnected items on both
rc4 boards; since lazy via checks it costs about 2x in time, not 4x.
The grid engine keeps 1, the old grid exactly.
A zone outline is where copper may go, not where it ends up: on the rc4
sense board the board-wide GND pours break into islands around other
nets' tracks, and treating every pad inside them as connected left 28
GND islands unjoined (63 unconnected against 42 without zones). Trust a
zone only when a no-tracks rule area covers it on its layer, as the
power stage's pours are covered; route the rest with tracks as before.
rc4, exact engine, resolution 2: power stage 5 unconnected (JAR 12),
sense board 41 (JAR 32), no DRC errors on either.
freeroute ignored the DSN's planes (KiCad's zones), so it never knew that
pads in a pour are joined by it. On the rc4 power stage those pads sit
inside "keep router off" areas as well, and every connection to them was
dropped without a word.
- Read (plane <net> <shape>) and put each zone on the board as a
ConductionArea of its net. Zones are not obstacles: KiCad refills them
around other nets' copper.
- The grid groups each net's pads and zone pieces that copper already
joins (a pad in a zone, two zones sharing a pad) and routes only between
groups. A connection may end on a zone by a via inside its copper, or by
a track where tracks are allowed.
- wire_keepout now blocks traces only, as KiCad's "no tracks" rule area
does; via_keepout only vias; keepout both, in the grid and in the exact
verifier.
rc4 power stage, exact engine, resolution 2: 2 unconnected, down from 8
(FreeRouting JAR: 12), no DRC errors.
Profiling the rc4 sense board at resolution 2 put half the time in
_via_placeable: every expanded cell checked a 5x5 via footprint on both
layers, for vias the search almost never takes. Queue a layer change as a
candidate and check its footprint (memoised per search) only when it is
popped. A check can only rule a via out, never make it cheaper, so this
cannot change which route is found.
The grid search estimated the remaining distance as a straight line even
when only orthogonal steps exist; Manhattan is the tighter estimate those
moves admit and still never overestimates. --heuristic-weight W (default
1) makes it weighted A*: routes at most W times the shortest (tested).
A cell was exactly one trace plus its clearance, so a trace owning its cell
was the whole clearance model, and the 0.4 mm grid could not reach most
pins of a 0.5 mm pitch package: between two 0.3 mm pads a trace's centre
has a legal band only 0.15 mm wide. Resolution K makes cells K times
finer and has every trace claim a halo, the cells whose centre is nearer
than trace + clearance to one of its own, released again on rip-up.
K = 1 has an empty halo and is the old grid exactly (tested). On a 0.5 mm
row, K = 4 gets every pin out where K = 1 gets two. On the rc4 boards, K = 2
cuts the sense board's opens from 66 to 41 (JAR: 32) and the power board's
to 7 (JAR: 12), still with no DRC errors from freeroute's own copper.
Every route starts and ends with a stub from the pin to its first grid
cell, and the grid never saw that copper: on the rc4 sense board 24 of 39
connections the exact verifier rejected had a later route pass beside an
earlier stub. A route's stubs now claim the cells they come near, a stub
must be clear of other nets to be used, and a route may start from any
enterable neighbour of the pin cell, costed by stub length (Manhattan on
the orthogonal track, so a pin on a cell centre still leaves straight).
Tests: a route near a rounded corner keeps its edge clearance; and a
strict xfail that pins down the remaining gap, a 0.5 mm pin row the grid
cannot give a column to every pin of.
- The exact engine verified copper per net and dropped a whole net on any
conflict: a 25-pin +3V3 net lost every connection to one bad segment.
Verify and accept per ratsnest connection instead.
- Pad and keepout cells were rounded outward, blocking cells up to half a
step clear of the copper; 62 of the 64 pins of a 0.5 mm pitch LQFP could
not start a route. Block only cells whose centre is inside the clearance
zone, and let a net walk along inside its own pad, which adds no copper.
- Routing was bounded by the outline's bounding box, so tracks cut across
rounded corners. Keep the DSN boundary as a polygon; the grid blocks cells
too close to it (scanline plus per-edge, precomputed for vias) and the
exact and room engines reject copper that is.
- --edge-clearance MM sets copper-to-edge clearance, e.g. KiCad's
min_copper_edge_clearance, which the DSN does not carry. Default 0.
On the TerraHawk rc4 power stage the exact engine now matches the
FreeRouting JAR: same open count, no DRC errors, with a 0.5 mm edge rule.
Measured against KiCad's own pads on the TerraHawk rc4 power stage:
- Back-side parts kept their pads on F.Cu (32 pads). Mirror the copper
layer for a back-side part.
- Pad outlines were copied unrotated (29 pads on rotated parts). Place the
outline through the same chain as the pin: pin rotation, back-side
mirror, part rotation. Oval pads, a path with a width, came out as a
zero-height sliver; pad them out by the aperture.
- Multi-shape padstacks used the first shape; use the largest.
- Footprint-level keepouts were parsed but never placed. KiCad exports a
mounting hole's plated ring that way, so the router could not see it.
- Vias were sized like a trace (0.25 mm square for a 0.6 mm via), and every
remaining short and clearance error was a via. Read the size from the
DSN's via padstack and use it in all three engines; the grid engine now
reserves the cells a via's copper reaches.
With these the exact and room engines route that board with no shorts,
clearance or other DRC errors. Pins now carry their component reference.
On the TerraHawk rc4 power stage every engine misread the board's rule areas:
- A no-track ring along the board edge arrives as one keyhole polygon, and
keepout polygons were wrapped in a single convex tile, so the ring became
the whole board. Split keepout polygons into convex pieces instead.
- The grid engine rasterized an area by its overall bounding box, which for
that ring is again the whole board; it routed nothing. Rasterize per tile.
- "Keep router off" rule areas arrive as wire_keepout, which the reader did
not know, so tracks ran through them (172 DRC items_not_allowed). Read it
at board and image level. For now it blocks vias too, which is stricter
than KiCad and cannot create a violation.
- place_keepout only restricts part placement; stop treating it as copper.
The oracle now runs the JAR headless; it was opening FreeRouting's GUI on
the desktop for every oracle test.
SES lengths and coordinates were written in the DSN's unit while the
session declared (resolution um 10), so readers that follow FreeRouting's
SesWriter (KiCad's importer, mckicad's ses_apply) placed every track at a
tenth of its size. The JAR, run on the same DSN, writes F1 at 65000 where
freeroute wrote 6500.
Scale every length by the resolution factor: placements, via padstack
shapes, wire widths and points, and vias. Rotations stay as they are. The
router tests compared SES output against DSN pad positions and so had the
same mistake built in; they now derive the expected values through
ses_scale.
Captures the three non-negotiable invariants (DRC-clean output, exact integer
geometry, determinism), the engine matrix, the reference/ GPL rule, the JAR
oracle strategy, and the lessons that cost real debugging: KiCad's /*52 net
names, exact clearance being load-bearing, shove not being a densifier, and the
falsifiable-fixture bar for any density claim.
Force-added past the global ignore: this is a private repo and the file is
excluded from the sdist, so it never reaches PyPI.
CalVer per project convention: the date communicates when the router was last
validated against the reference FreeRouting engine and real KiCad output. The
Alpha classifier and README carry the maturity signal.
Add the GPLv3 text (the package declared the licence but never shipped it),
a py.typed marker, and an sdist include/exclude allowlist so the distribution
carries only src/, docs/, README, LICENSE and pyproject. reference/ is a clone
of the GPL FreeRouting Java tree kept purely as a porting reference and must
never be redistributed inside this package; tests, caches and build output are
excluded too.
Switch to the PEP 639 SPDX licence expression with license-files, widen the
classifiers, and add an -o alias for -do. Rewrite the README around the engine
matrix: what each track and pass actually does, and what this is not (no
FreeRouting density parity on dense boards, and --diagonal is a recovery and
shortening pass, not a diagonal-native search).
The grid MVP was the only track reachable from the command line, so the
continuous room engine, channel packing, shove, and 45-degree routing were
unusable by anything shelling out to freeroute.
Add --engine {grid,exact,room} (default grid, unchanged behaviour) plus
--pack / --shove / --diagonal, and honour -mp on the tracks that implement
rip-up passes. Not every engine supports every pass; ENGINE_OPTIONS is the
matrix and an unsupported combination exits 2 with a message naming the
engines that do support the flag, rather than silently no-opping.
Add diagonal_win.dsn: a 45-degree pin-to-pin net with a keepout placed inside the
diagonal's bounding box but clear of the true thin copper. Orthogonal routing
connects it with a Manhattan-length staircase (1600000); diagonal routing
replaces that with a single 45-degree trace at the ideal hypotenuse length
(1131371) -- a 29% reduction -- DRC-clean via the search tree and via an
independent reconstruction of the emitted 45-degree copper's clearance octagon.
test_exact_clearance_is_load_bearing shows the keepout overlaps the coarse
bounding box but not the exact octagon, so the exact clearance (not the box) is
what lets the diagonal route. Determinism, endpoints-on-pads, valid SES, and
diagonal-off-equals-orthogonal (additive, over four fixtures) are covered.
Add diagonal=True to the exact router: after orthogonal routing, each 2-pin net
is retried as the shortest exactly-clean octilinear (0/45/90/135-degree) trace.
A dropped net is recovered; an orthogonal route is replaced only when the
diagonal is strictly shorter. Candidates are the direct 45-degree segment or the
two diagonal-plus-axis two-benders; each segment's diagonal copper is covered by
the exact octagon and its clearance checked with clearance_conflict_shape, so a
trace fits a diagonal corridor the bounding-box cover would reject.
Endpoints stay on the pads; diagonal copper is stored with its exact octagon so
later nets clear it precisely; same-net copper is ignored during the check so the
candidate is verified against every other net before the orthogonal copper is
ripped. With diagonal=False the pass is skipped and output is byte-for-byte the
orthogonal router. Threaded through route_dsn_board_exact /
build_exact_routing_result / route(engine="exact").
Add optional exact/seg fields to TreeShape for a diagonal trace tile: tile holds
the bounding box for broad-phase bucketing, exact holds the tight octagon copper,
and seg=(a,b,half_width) rebuilds the clearance region on demand. Every
orthogonal tile leaves both None and takes the identical box code path, so the
box-only tree stays byte-for-byte unchanged (whole suite green).
clearance_conflict now tests a stored diagonal tile with an exact 2-D overlap;
clearance_conflict_shape checks a diagonal trace's clearance octagon against
stored copper; has_violation runs the exact octagon test whenever either tile is
diagonal and the plain box overlap otherwise. All exact-integer via Simplex
intersection.
A diagonal segment's copper has no axis-aligned exact tile, so segment_box falls
back to the bounding box -- a huge conservative cover that makes diagonal traces
conflict with everything. segment_octagon builds the tight cover instead: the
copper of a 0/45/90/135-degree segment swept by a radius, as a convex integer
Simplex bounded by the four axis and four x+y / x-y diagonal half-planes.
Diagonal bounds widen by ceil_sqrt2(radius) -- the exact integer ceiling of
radius*sqrt(2) via math.isqrt, no float -- rounded outward, so the octagon is a
provable superset of segment (+) disk(radius) (verified over 20k sampled copper
points). Growing by clearance is just a larger radius, giving both the stored
copper tile and the queried clearance region. overlaps_2d tests a real 2-D
overlap via exact Simplex intersection, accepting mixed IntBox/Simplex tiles.
Add channel_pack.dsn: three nets whose pads spread wider than a wall gap, so they
must converge and pack into lanes. Plain room routing (shove on or off) drops >= 1
net under every net ordering; the packer connects all three under every ordering,
DRC-clean (search tree and independent emitted-copper check).
The falsifiable boundary test sweeps the gap width for N=3 and N=4 and asserts the
packer connects all N iff the gap meets the geometric feasibility width
(N-1)*(width+clearance) + 2*(half_width+clearance), never routes fewer than greedy
below it, and never false-packs. Determinism, endpoints-on-pads, valid SES, and
pack=False-equals-default (additive) are covered.
Greedy per-net room routing sends each net through a shared obstacle gap on its
own independently chosen path; the paths collide and some nets drop even when the
gap is physically wide enough for all of them. shove cannot fix this -- it
relocates blockers into existing space, it does not pack lanes.
Add an opt-in pre-pass (pack=True, independent of shove) that groups 2-pin nets
which must cross a common gap, assigns each a parallel lane across the gap's
usable width (pad-ordered so lanes never cross, spaced >= width + clearance) with
staggered fan-in trunks so the converge/diverge jogs never overlap, and realizes
each as an orthogonal trace verified exactly DRC-clean. Packing is all-or-nothing
per group, so it never leaves a channel worse than greedy. With pack=False the
pre-pass is skipped and the output is unchanged.
Scope: single-layer 2-pin nets through one axis-aligned gap, with enough fan
depth to stack the trunks. Threaded through route_dsn_board_rooms /
build_rooms_routing_result / route(engine="room").
Add true_density.dsn: NET_C is a full-height wall on the sole signal layer,
so plain room routing drops a net under every net ordering (an order-independent
obstruction, verified over all permutations), unlike rooms_shove_channel whose
plain drop fits under a reorder.
Test that shove recovers nothing on this board (shoves == 0, same routed set as
plain) while staying DRC-clean, and that the only difference from the recoverable
shove_channel is the wall's span. shove relocates a blocker into existing free
space; where the wall leaves no headroom there is nowhere to relocate to, so it
cannot manufacture density. The recover/fail boundary is the free-space boundary.
_shove_recover_room clobbered a net's _Accepted with a fresh one whose
wire_owners held only the just-recovered wire, while its wires list (the
shared result.wires[net_no]) already held earlier committed connections.
That left wire_owners shorter than the parallel wires list, so a later
net's shove either removed the wrong wire's copper from the search tree
(a silent different-net short: copper stayed in the emitted SES but
vanished from the tree, and has_violation falsely certified it clean) or
raised IndexError and aborted the whole route.
Reuse the net's existing _Accepted on recovery so wire_owners stays
parallel to wires. Fixes both the silent DRC violation and the crash
sibling without touching the DRC-clean or determinism invariants.
Also correct the overstated "density win" framing of the shove_channel
fixture: its plain-routing net drop is an artifact of greedy net order,
not a density limit (the same board routes both nets with zero shoves
under a reorder or with the cosmetic seals removed). Reframed the test
and docstring; the genuine sub-cell win remains the wide_door occupancy
packing (shoves == 0).
Regression tests assert the invariant against the emitted output
directly (reconstructed copper tiles), not just the tree that the bug
fooled, plus determinism, no-crash, and the ordering-artifact bound.
rooms_shove_channel is the headline density win: a sealed NET_A wall
leaves no top<->bottom door, so plain room routing drops NET_C; with
shove=True NET_A is nudged aside and both nets route, DRC-clean and
deterministic. rooms_wide_door seats two nets through one wide gap at
distinct projected gate points with no shove.
Cover the occupancy-aware projection math directly, assert the four
foundation fixtures stay byte-identical under shove=True (shoves==0), and
extend the room DRC-clean parametrize with both new fixtures.
Both upgrades sit behind an opt-in shove=False flag, so the default code
path is byte-for-byte the foundation router (existing room tests are
unchanged by construction).
With shove=True:
- Gate placement becomes occupancy-aware: each edge gate is projected
into the largest other-net-free sub-span of the door, with the reserved
intervals read straight from the search tree. With margin = clearance +
half_width the room decomposition already keeps doors clearance-clear,
so this reduces exactly to the old midpoint clamp where no other-net
copper crosses the door.
- When a net would be dropped, a transactional shove-in-rooms recovery
nudges the committed trace(s) blocking it aside with the exact shove
primitive and re-verifies against the clearance oracle, so both nets fit
through one channel. Every moved trace is journaled and rolled back to
its exact original tiles and wires on any failure -- the room engine
mutates already-committed nets, so the rollback is real rather than the
exact track's no-op.
The board-wide has_violation(clearance) gate stays the final assertion.
Thread the flag through pipeline.route_dsn_board_rooms /
build_rooms_routing_result / route(engine='room', shove=...).
Factor the exact router's shove geometry into route/shove.py: the
conflicting-segment finder and displacement ladder are re-exported from
exact_router (kept byte-unchanged so its shove count and geometry stay
identical), and a room-oriented trace placement helper is added that
returns the inserted tile owner ids for a journaled rollback.
Adds narrow_channel.dsn: a keepout wall with a narrow gap positioned off the
routing grid, so no grid cell centre falls inside the gap's free centreline.
The exact/grid track drops the net (cell quantization can't fit a centreline
in the channel); the continuous room track routes it. This is the headline
result that continuous free-space beats the grid.
Tests: the room router's output is DRC-clean (exact has_violation is None) on
simple/crossing/kicad/narrow_channel; it connects simple/crossing (with a via)
and all four multi-pin KiCad nets; traces are orthogonal and end on pads; and
the continuity contrast asserts exact routes 0 while rooms routes 1, both
clean. Oracle-gated connectivity parity on simple/crossing.
A third engine (engine='room') that replaces the fixed routing grid with
FreeRouting's free-space rooms. Per layer, the space not occupied by items
(inflated by the routing clearance) is decomposed into exact rectangles via
coordinate compression (the orthogonal analogue of
CompleteFreeSpaceExpansionRoom); adjacent rooms share an edge door, and a via
door joins overlapping rooms on two layers. The maze search is an A* over
rooms-through-doors (cost = geometric distance + via cost), so a path may
cross a door anywhere along its width rather than at a quantized cell. The
found room sequence is realized into an exact orthogonal Polyline through the
door mid-gates and verified DRC-clean against the ShapeSearchTree, dropping
the net if it cannot be made exactly clean.
Maps to autoroute/: ExpansionRoom / CompleteFreeSpaceExpansionRoom (rooms),
ExpansionDoor / TargetItemExpansionDoor (doors), MazeSearchAlgo /
AutorouteEngine (search), LocateFoundConnectionAlgo (realization).
Foundation scope: rooms are axis-aligned rectangles (not general Simplex
tiles); realization uses an L-connector through door mid-gates. Any-angle
rooms, optimal gate placement (true sub-cell shove), and optimization passes
are follow-ups. The grid and exact tracks are unchanged; this is opt-in.
Routes simple/crossing (with a via) and all four multi-pin nets of the real
KiCad board, DRC-clean.
Adds shove_needed.dsn: one signal layer (no vias), where NET_A routes as
a straight wall that boxes NET_C out under greedy ordering. Tests assert
the contrast: without shove NET_C is dropped (result still DRC-clean, just
incomplete); with shove NET_A is moved aside and both nets connect, the
exact has_violation check is clean, endpoints stay on pads, and the result
is deterministic. Regression tests confirm shove is a no-op on the simple/
crossing/ripup/kicad boards (same connectivity, DRC-clean, zero shoves).
Oracle-gated parity is included (skips when the JAR cannot route the
synthetic single-layer board).
Before a dropped net is abandoned, the exact router (with shove=True)
tries to make room by moving an existing trace aside instead: it forms a
straight orthogonal candidate for the dropped net, finds the axis-aligned
trace segments that cross it, and shoves each one perpendicular (via
shove_segment) far enough to restore clearance. A shove is accepted only
if the moved trace still connects its pads, clears every other item
exactly against the ShapeSearchTree, and stays inside the board outline,
so DRC-cleanliness is preserved by construction. Shoves are bounded
(max_shove_depth, per-trace shove_cap) and deterministic. The dropped set
now includes nets the grid failed to route, not only exact-clearance
rejections, so shove can recover them.
The pipeline exposes shove (and rip_up) through route_dsn_board_exact /
build_exact_routing_result / route(engine='exact', shove=True); default
off, so the grid track and the no-shove exact track are unchanged.
Displaces one segment of a trace polyline perpendicular by a given
offset, keeping the trace orthogonal and its two pad endpoints fixed: a
jog corner is inserted when the moved segment is first or last, and
interior neighbours simply resize. Collinear/duplicate corners are
dropped. This is the foundation for shoving a trace aside to make room
for another. Unit-tested for single-segment, interior-segment, and
orthogonality-preservation cases.
Adds route/exact_router.py: routes with the grid track's rip-up/multi-layer/
via topology search in orthogonal mode (axis-aligned segments -> exact IntBox
copper), then verifies clearance exactly against a ShapeSearchTree. A net
whose copper would come closer than the clearance to any accepted item is
dropped, so the emitted geometry is DRC-clean by construction; endpoints stay
exactly on pads and vias at real layer transitions.
GridRouter gains a non-breaking orthogonal option (4-connected). The pipeline
gains an engine= selector: 'grid' (default, unchanged, highest coverage) or
'exact' (DRC-verified). The grid MVP stays the fallback.
On the designed boards the exact track routes every net DRC-clean; on the
real KiCad board it matches the grid's connectivity (4/4 multi-pin nets) and
is DRC-clean where the grid MVP's widened centrelines are not. Deferred:
shove, and 45-degree/IntOctagon trace caps.
Tests: the exact router's output has no clearance violation (exact IntBox
check) on every fixture; endpoints on pads; traces orthogonal; crossing uses
a via; and a direct comparison showing the exact track is clean where the
grid MVP violates on the real board. Oracle-gated parity for engine='exact'.
A bounding-box spatial hash (broad phase) with exact IntBox-intersection
narrow phase, replacing the exact router's occupancy grid. Supports
insert/remove-owner and overlap/region queries, a clearance_conflict test
(strict 2-D overlap of the clearance-expanded box, so tiles exactly a
clearance apart are allowed), and has_violation for the whole-board DRC
check. Tiles are tagged routed vs static so pad-vs-pad spacing in the
source design is not counted as a routing violation.
Ports the trace-relevant subset of geometry/planar/{Polyline,PolylineShape,
LineSegment}: a Polyline is a trace centreline (integer corners); a
PolylineShape is that centreline swept by the half-width, returned as exact
convex IntBox tiles (one flush box per orthogonal segment plus a square at
each interior corner to cover the turn). Exact for orthogonal segments; a
conservative bounding box for a diagonal one (never under-reports an
overlap). 45-degree/IntOctagon caps are deferred.
The pyproject declared a freeroute script but freeroute.cli was missing, so the
entry point was broken. Add it: reads a Specctra .dsn, routes (rip-up on by
default), writes SES. Accepts FreeRouting-compatible -de/-do flags so it can be
dropped in wherever 'java -jar freerouting.jar' was called.
Adds ripup_needed.dsn: one signal layer (B.Cu is a power plane, so vias
are impossible) with two nets whose greedy order strands NET_B, but where
rip-up reroutes NET_A around NET_B to connect both.
Tests: greedy (rip_up=False) leaves >= 1 connection unrouted with no vias;
rip-up connects all with no vias; the result is deterministic across
re-runs; rip-up keeps endpoints on pads and produces no same-layer cross
between nets (ripped traces leave no orphaned occupancy). The crossing
board's single-layer test now targets the greedy path explicitly, and a
regression test asserts the real KiCad board still connects all four of
its multi-pin nets. Oracle-gated parity added for the congested board.
Restructures the router so occupancy is tracked per connection, then adds
a rip-up-and-retry loop that recovers from bad greedy net orderings.
When a connection cannot reach its target through free space, a rip-up
search may pass through other nets' traces at an escalating penalty; the
router rips up the connections that path crosses (fully removing their
occupancy), routes the failing connection, and re-queues the ripped
connections. Passes iterate up to max_passes, keeping the best
(fewest-unrouted) result and stopping on full success or when a pass
changes nothing.
Thrash prevention: each connection may be ripped at most rip_cap times,
and the rip penalty escalates with a connection's rip count, so
repeatedly-ripped connections harden into walls. Fully deterministic
(sorted rip sets, connection-creation order, tuple-keyed A* — no RNG).
route()/route_dsn_board()/route_board() take rip_up and max_passes;
rip_up=False reproduces the previous greedy single-pass behaviour for
comparison. RouteResult now reports the unrouted connection count.
Adds crossing_2net.dsn: NET_A is a full-width horizontal wall and NET_B's
pins sit above and below it, so single-layer routing cannot connect NET_B
but two-layer routing succeeds by dipping to the other layer through vias.
Tests assert the multi-layer invariants: single-layer routing leaves a net
unrouted with no vias; multi-layer connects both nets with >= 1 via; the
SES contains a via scope; every via sits at a real layer transition of its
net's path; trace/via endpoints still land on the pads; traces stay on
valid layers and within the outline; and no two different nets cross on the
same layer. Oracle-gated parity tests cover the simple and crossing boards.
rsp2k