ribbon: per-scaffold rotation optimizer + orientation marks (#127) - #128
Open
conchoecia wants to merge 15 commits into
Open
ribbon: per-scaffold rotation optimizer + orientation marks (#127)#128conchoecia wants to merge 15 commits into
conchoecia wants to merge 15 commits into
Conversation
Scaffold plot direction is now chosen independently of the assembly
fasta direction so that the number of bezier-line crossings between each
adjacent species pair is minimized. Cascades top-down: the top species
is pinned fasta-forward, each species below picks its flips with the
species above already decided. Inversion count between the top-x order
and the bottom-x order is used as the crossing score (paths are
monotonic in y, so two paths cross iff their endpoint orderings flip).
A per-chrom independent seed (cheap) is followed by an iterated
flip-greedy refinement on the global crossing count.
Each chromosome bar gets a small filled triangle at the fasta-3' end so
the viewer can read plot-direction vs fasta-direction at a glance.
Two new config keys, both default True, gate the new behavior:
optimize_chrom_rotation: True
show_orientation_marks: True
Setting both to False reproduces the previous plot byte-for-byte.
Also:
- scripts/odp_rbh_to_ribbon used to carry its own inline copy of
ribbon_plot/_quality_check_chromosome_list/_optimize_spA_based_on_rbh
/plot_bezier_lines that had drifted out of sync with
scripts/odp_ribbon_plot.py. The snakefile now imports the module
directly so there is one source of truth.
- Fixed two pre-existing bugs surfaced while building the
Lepidoptera test case:
* groupby(...).apply(...) + index.names assignment broke on
pandas >=2; replaced with an equivalent cumcount-based
construction of the within-scaffold gene index.
* When plot_all=True the "alpha" column was selected without
having been populated; now it is computed in both branches.
Closes #127.
Follow-up to the previous commit on this branch, addressing review feedback on #127. Top-row ordering ---------------- optimal-size used plain value_counts() to seed the very first row, which ignores which row-2 chromosome each top-row scaffold actually anchors to. Two scaffolds with the same best-FET partner in row 2 could end up scattered across the top row, forcing crossings the cascade below could not undo. The new seed _seed_top_order_from_partner groups top-row scaffolds by their best-FET partner in row 2, orders partner groups by total shared ortholog count, and within each group orders members by FET strength. With this, both Plutella xylostella adjacencies the user called out (CM/509.1 next to CM/495.1, CM/506.1 next to CM/501.1; both pairs share a row-2 best partner) now hold. Additional chr_sort_order options --------------------------------- For comparison and for highly rearranged genomes where the existing FET-best-partner heuristic struggles, four new modes are now selectable: optimal-barycenter single top-down barycenter pass optimal-barycenter-iter iterated barycenter sweeps to fixed point optimal-median Eades-style median heuristic optimal-swap adjacent-swap crossing-count polish All take the same chr_sort_order config slot. The snakefile validator was extended to accept them. On the Lep test case the existing optimal-size still wins; the new modes are meant for datasets with no single best partner per chrom. Orientation marks and labels ---------------------------- The orientation glyph at the fasta-3' end of each chrom bar is now a chevron tip on the bar itself (head_h=0.07, head_len_cap=0.0045) rather than a separate triangle, so it stays subtle on dense plots. Scaffold labels are now plotted next to each chrom bar at 90 degrees CCW, anchored just left-and-above the bar's left end, and shortened to "<first 2 chars>/<last 3 of stem>.<version>" — so NC_051358.1 is labelled NC/358.1 rather than the full id. Short ids pass through unchanged. The plot ylim was widened to fit the rotated labels. Closes none new; continues #127.
A row-by-row cascade -- even bidirectional with crossing-aware back-tracking -- only sees one neighbor at a time, so a row-4 swap that needs to propagate up to row 0 stays invisible to it. odp #127 review surfaced concrete failures: CM/509.1 next to CM/495.1, CM/506.1 next to CM/501.1, CM/522.1 next to CM/500.1, and CM/505.1 / NC/729.1 on the other end of CM/523.1 all need the algorithm to look at the whole stack at once, not pairs. New chr_sort_order: optimal-lg ------------------------------- Builds a chromosome-level graph from every adjacent-pair RBH and partitions chromosomes into synteny linkage components. Each species' chromosomes inherit a cluster id; the row ordering uses cluster id as the primary sort key and the existing FET-best-partner cascade as the within-cluster tie-break. Members of one LG share an x-region across every row, so ribbons run straight down within an LG. Graph construction: - edges kept only when both sides' FET-strongest partner (top-k=1) is the other -- "strict mutual best." Pure union of top-1 picks or top-k>=2 collapses the graph to one giant component on Lep-style data with universal residual homology. - global orphan rescue: a chromosome that received no mutual-best edge anywhere in the graph gets exactly one one-way edge to its FET-strongest partner across all adjacent-pair RBHs. Restricting the rescue to nodes with no other edges prevents a chrom that already lives in an LG component on one side from accidentally bridging two distinct LG components on the other side via a fusion hub. - cross-cluster edge weights are computed from all FET-significant pairs (not just mutual best); the cluster ordering is a greedy nearest-neighbor traversal weighted by those cross weights so that fusion partners (e.g. NC/714.1's cluster and NC/729.1's cluster sharing Bombyx NC_051368.1) end up plot-adjacent. Within each cluster, _seed_top_order_within and _cascade_within run the same FET-best-partner heuristic the existing cascade uses, but restricted to that cluster's chromosomes, so siblings sharing a downstream chrom stay adjacent. Other changes ------------- - chr_sort_order: "optimal-lg" added to the snakefile validator. - The chrom scaffold label is now compacted to "<first 2 chars>/<3 chars before .N>.N" form (NC_051358.1 -> NC/358.1) so the trailing decimal suffix stays visible. Labels nudged slightly off the bar's left end and rotated CCW 90 degrees to read bottom-up next to the chrom. - Orientation arrow shrunk again (head_h=0.07, head_len_cap=0.0045) so the glyph is a subtle chevron tip on the bar instead of a tall separate triangle. Continues #127.
The previous commits were a chain of heuristics, each hoping that a sensible-looking re-ordering would happen to reduce ribbon crossings. None of them ever scored against the true crossing count, so each left a different residual set of crossings that the next commit then had to chase. This commit adds the actual objective and a direct search against it. What's new ---------- - _fenwick_weighted_inversion: O(n log n) weighted inversion count over (top_x, bot_x) line endpoints, the exact bezier-crossing count between an adjacent species pair. - _build_pair_line_data: per-pair record cache so the search loop never re-touches the RBH dataframe. - _score_pair_lines: evaluates one pair given current chrom orders and per-chrom flips. - _crossing_local_search: greedy improvement-monotone refinement on the global weighted score. Seed = whatever optimal-lg + the flip cascade produced (do not throw the cascade output away; refine it). Moves per chromosome: F/R flip, swap with adjacent chrom in row. Species traversal order per pass: most-rearranged first, then outward. Any move that reduces the total score is kept; iteration stops when a full pass produces no improvement. FET-significant orthologs are weighted 1000x faint ones, per the user's request -- the search effectively minimizes crossings of the bold ribbons first, only using the faint ones as tiebreak. Runtime on the 12-species Lepidoptera test: ~3 min for 15 passes. The seed score (which the previous LG ordering produced) was already strong, so the marginal drop is small (~4-5%), but the search now sees and acts on the actual objective rather than hoping a heuristic happens to produce it. Continues #127.
Two cheap speedups to the crossing-count local search added in the previous commit: - _build_pair_anchor_data: every distinct (top_chrom, bot_chrom) pair is collapsed to one record at the within-chrom mean rank on each side, weight = (fet_weight if best whole_FET <= 0.05 else 1) * shared-ortholog count. Reduces lines/pair from ~2000 to ~200 on Lep-style data. The chrom-flip and chrom-swap moves the search considers never touch within-chrom gene ordering anyway, so the per-ortholog score's extra resolution is wasted there; the anchor score sees the same delta-direction for every move the search could make. - early_stop_frac (default 0.05%): a full pass that improves the score by less than this fraction stops the loop. The long tail of monotone but visually-invisible improvements is skipped without changing the converged-on layout. max_passes raised from 15 to 50 (cap with early-stop, not at the working point). Lep test: 198 s, 15 passes, -4.5% -> 51 s, 11 passes, -6.8%. Roughly 4x faster and a deeper minimum (because the search now runs to actual convergence inside the cap). Continues #127.
User feedback on the previous commit: rotation cleanup at the end matters, and the coarse-only search regressed in places where the per-ortholog detail mattered. Changes ------- - Local search call site now runs two stages: a coarse-scoring stage for fast convergence to the right neighborhood, then a fine (per-ortholog) stage to recover the within-chrom detail the coarse anchors flattened out. - Move set extended: chromosome swaps now run at offsets 1, 2, 3 and 5 (was offset 1 only). Adjacent-only swap can't escape the local minimum where a chromosome's right neighbor is unrelated but its true partner sits two slots away. - After each stage's swap+flip loop converges, a dedicated final flip-only sweep visits every chromosome once more. By that point the ordering context for each scaffold's L/R orientation is fully determined, so some flips that weren't useful earlier in the search may now reduce the score. User called this out as key. Lep test runtime: 2:35 for both stages + final sweeps (vs 51 s single-stage coarse / 3:18 single-stage fine). Stage 1 ends at score 996e9 (coarse metric); stage 2 starts the per-ortholog metric at 1.38e12 and ends at 1.16e12. Continues #127.
User feedback on the previous commit:
- y-axis species labels were clipped to ~6 characters at the left.
- A chromosome that crosses many others above it should get special
attention to move it.
- Headroom remains on rearranged-species pairs; longer-range moves
should be tried.
Changes
-------
- Chromosome-level priority: within each species visit, chroms are
now visited worst-first by the weight of FET-significant lines
incident to them. The single chromosome that crosses many others
(the user's "one above that crosses many") is the first one the
search tries to move.
- Insertion move: each chrom can be inserted at any other slot in
the row (not just swapped with a fixed neighbor). A chrom whose
true partner sits five slots away can hop straight there even when
every individual adjacent swap raises the score because the route
passes through unrelated chroms.
- y-axis species label fix:
* caller can pass `species_labels` dict (the run_ribbon driver now
pulls it from `config["tree"]["tip_label_map"]` if present);
* default fallback strips the "-taxid-GCAxxx" / "-taxid-GCFxxx"
accession suffix;
* figure widened from 8" to 10" and panel layout adjusted so a full
"Plutella xylostella"-style label fits in the left margin instead
of being clipped to its rightmost few characters.
Lep test runtime: ~10 min. Coarse stage 1069B->947B (-11.4%), fine
1317B->1111B. Score and visual quality both improved on the same
seed -- insertion + priority found moves the swap-only search missed.
Continues #127.
User feedback: the previous wide figure dropped the sample accession and felt too wide. The plot is now a single-column 180 mm (7.087") wide and the y-axis labels keep the assembly accession on a second line below the species name: Plutella xylostella GCA019096205.1 The species_labels dict (typically threaded from config["tree"]["tip_label_map"]) supplies the human-friendly first line; the second line is parsed from the species id's trailing GC[AF]xxx.x token. If no tip_label_map is supplied we fall back to stripping the "-taxid-GCxxx" suffix as the primary label. Panel width: 4.6", left margin 2.187". Fits two-line labels at fontsize 7 without clipping. Continues #127.
… title
User feedback on the previous commit:
- panel title sat too far above the plot
- species names should be italicized
- accession number should be a slightly greyer color
- labels should be robust to missing inputs / fall back gracefully
- there was too much negative space to the left of the labels
- the search was getting stuck because single-row moves
inevitably crossed something else; some chrom pairs need to
be pulled together across rows ("sticky" moves)
Changes
-------
- Panel titles now sit 0.03" above the plot (was 0.25") and use
fontsize 10 (was 12). Right over the panel like a header.
- y-axis labels redrawn as a pair of manual Text() artists per row:
species name italic at fontsize 7, assembly accession in #888888
grey at fontsize 6 below. Matplotlib tick labels cannot mix
styles in a single string, so this is the cleanest way.
- Robustness: species_labels dict missing or partially populated is
handled (falls back to the stripped genus+species concatenation
off the species id, then to the bare species id if neither is
available). Accession is omitted from the label if the species id
doesn't contain a GC[AF]xxx.x token.
- Panel widened to 5.4" (was 4.6") inside the 7.087" / 180 mm
figure, tightening the left whitespace.
- Sticky LG-block move: each chromosome carries an FET-graph
cluster id (from _build_fet_lg_clusters). The new move type, run
inside the local search after the per-chrom flip / swap / insert
moves, tries dragging every member of a cluster to a candidate
slot in its own row simultaneously. Lets the optimizer escape
the local minimum where a single-row move would have to cross
other things to reach its true partner column -- exactly the
"guided annealing" the user described. The candidate slot set
is the current slot of any cluster member in any row, so the
block move always lands somewhere already occupied by one of
its own members.
Lep test runtime: 17:08 (was 9:43 without sticky block). Score
1069B -> 865B in the coarse stage (-19%, vs -11.4% prior) and
1211B -> 1022B in the fine stage; the block move clearly finds
moves the per-row search misses.
Continues #127.
Major reworking of the chromosome ordering / flip optimization
pipeline. Replaces the prior local-search-only approach with a
modified-Sugiyama brushing sweep seeded by optimal-lg, then iterated
random restarts on top.
Pipeline (when optimize_chrom_rotation=True):
optimal-lg cluster seed
-> flip cascade (top-down, then a final all-rows greedy pass
including row 0)
-> brushing sweep: alternating top-down and bottom-up FET-
weighted barycenter, with the flip cascade re-run after each
direction; per-row "feedback" greedy moves on the top row
(after td) and bottom row (after bu); best-seen state
tracked across all sweeps; patience-based convergence
-> iterated restart: small random perturbation of the best
state, brushing again, keep global best across restarts
-> render
What's new in detail
--------------------
_brushing_sweep: per-generation td + tm + bu + bm. Each direction
runs a FET-weighted barycenter pass over every chromosome of every
row, then a flip cascade. Top-row / bottom-row feedback (tm / bm)
is a tension-prioritised greedy on the row that the direction
just settled, accepting moves only if the global crossing count
drops. Best-seen state is restored at the end, so a transient
bu regression does not stick.
_optimize_chrom_flips_top_down: now accepts an `initial_flip`
argument so callers can preserve row-0 (and row-1+) flips chosen
by upstream moves -- previously this function reset row 0 to all-
False every call and undid any row-0 flip accepted by
try_top_moves. Adds a final boundary pass that revisits every
chromosome in every row (including row 0) with a full-crossing-
count greedy, catching flips the row-by-row cascade missed.
_render_ribbon_figure: extracted from ribbon_plot so the brushing
loop can snapshot a PDF after every sweep ("brush stroke"), with
filenames sorted to match run order:
{base}_brush_gen{N}_{01_td|02_tm|03_bu|04_bm}.pdf
ribbon_plot:
- Iterated random restart wrapper. After the initial brushing
descent, the best state is perturbed (random shuffle of a
small window of chromosomes in a few randomly chosen rows) and
brushing is re-run. Keeps the global best across all restarts.
- Stale snapshot PDFs from prior runs are deleted at the start
so the workspace only shows the current run's progression.
- flush=True on all log prints so SLURM's block buffering does
not hide progress on long runs.
Plot layout:
- figWidth = 7.087" (single-column journal 180 mm).
- panelWidth widened to 6.3" (tighter left/right margins) so
more of the panel is the data.
- Panel titles raised to +0.8" above the panel top so the
rotated scaffold labels at the top of each panel no longer
bleed into the title.
- Two-line y-axis labels: italic species name (fontsize 7) over
grey accession in #888888 (fontsize 6), drawn as paired
Text() artists so matplotlib's tick-label-can't-mix-styles
limitation does not apply. Robust to missing tip_label_map.
Lepidoptera 12-species test, score 1069B -> 374B (-65%).
Single-descent run ~3 min; full 8-restart pipeline ~30 min without
per-gen snapshots, ~70 min with snapshots.
Notes:
docs/rbh-to-ribbon-sorting-notes.md captures the design log,
what worked, what didn't, and open problems.
Continues #127.
Two fixes prompted by user inspection of a converged plot where
CM/509.1 and CM/506.1 in Plutella xylostella were left in suboptimal
orientation despite the boundary flip pass.
Root cause
----------
The brushing sweep tracks best-seen state by an anchor-coarse score
(one record per (top_chrom, bot_chrom) pair at within-chrom mean
rank, for speed). When best_flips is captured, the flips are
optimal w.r.t. that coarsened metric. The per-ortholog FET-weighted
score the boundary flip pass uses can disagree -- and does, by
~3.7 billion crossings on CM/509.1's flip alone on the Lep test.
After brushing restored best_orders/best_flips it did NOT re-run
the flip cascade on the restored state, so the suboptimal flips
captured at best_score moment stuck around.
Fix
---
- _brushing_sweep: after restoring best_orders/best_flips, call
_optimize_chrom_flips_top_down one more time with
initial_flip=current state. The cascade's boundary pass is per-
ortholog FET-weighted so it catches the flips the coarse-score
brushing missed.
- ribbon_plot's iterated-restart wrapper: same final cascade after
restoring global_best_state.
- ribbon_plot: after writing the final PDF, persist
{species_order, chromorder, chromflip} as JSON next to it.
Future diagnostics (why isn't chrom X flipped?) can load this
and inspect/score the optimization output without re-running
the 30+ minute pipeline.
Verified by running the standalone cascade on the previous run's
saved state JSON:
Before: CM/509.1=R CM/506.1=F CM/495.1=R CM/501.1=F
After: CM/509.1=F CM/506.1=R CM/495.1=R CM/501.1=F
matching user-identified misses exactly. CM/495/501 correctly
left unchanged -- per-ortholog says their current orientation is
optimal.
panelWidth reduced 6.0 -> 5.6 so two-line italic species labels
no longer clip past the figure left edge at panelWidth=6.0.
Continues #127.
User asked for <10 min, ideally <2 min on a 12-species Lep test that
previously took 6.5h for thorough mode. This commit gets fast to
~3 min and medium to ~5 min: 119x and 76x speedups respectively.
What was done
-------------
1. Replaced the pandas-.apply hot loop in _eval_pair with a numpy-
vectorized fast path that scores against precomputed per-pair
numpy arrays. Tables are built once per cascade call instead
of being reconstructed for every flip attempt.
2. Added integer chromosome indices to pair tables. Position lookup
per line is now one numpy fancy-index instead of a per-element
pandas Series.map. cProfile showed pandas.Series.map at 1612s
cumulative on the unprofiled-9-min fast run; this change makes
the lookup essentially free.
3. JIT-compiled the Fenwick weighted-inversion-count inner loop with
numba (pure-Python fallback if numba is missing). The inversion
count was the next bottleneck after the indexing fix.
4. Routed the legacy _count_inversions through the same JIT'd
Fenwick when numba is available.
5. Added an optimization_level config key with values
fast - no restarts, ~3 min, -63% from seed
medium - 2 restarts, ~5 min, -66% (default)
slow - 5 restarts, ~10-15 min, -66%
thorough - 8 restarts + per-gen snapshot PDFs, ~hour, -66%
so users can dial cost/quality. Snapshots are gated to thorough
only (they roughly double wall time at any level).
Lep 12-species test before/after:
thorough 6h35m -> ~hour (snapshots dominate now)
medium 18 m -> 5:14
fast 9 m -> 3:19
Score parity (purely computational, not algorithmic):
fast quality unchanged at 389B (-63.4%)
medium quality unchanged at 363B (-65.9%)
slow/thorough plateau at 362B
Continues #127.
Continues from the previous commit's 120x speedup. Remaining medium runtime was dominated by leftover pandas.Series.map calls in the brushing-coarse score path and a per-row lambda inside the flip cascade's bot_positions, both of which only showed up in profile after the per-ortholog hot path was vectorized. Changes ------- 1. _build_pair_anchor_data now emits the same integer-chrom-index arrays _build_pair_line_table does. _score_pair_lines takes its fast path on coarse data too, dropping pandas.Series.map from brushing's total_crossings hot loop. 2. bot_positions inside _optimize_chrom_flips_top_down precomputes the bot_scaf integer index once and uses fancy-indexed numpy lookup instead of a pd.Series(...).map(lambda) call. Profile showed ~85s of that lambda on the Lep test. 3. barycenter_position in _brushing_sweep uses np.where on a pre-built flip mask column instead of a .apply(lambda) row-by-row. 4. _optimize_chrom_flips_top_down's inner cascade calls inside brushing now take max_passes=3 (was 8 default). On Lep the inner cascade converges in 1-2 passes anyway; this skips the final 5 stagnant ones. 5. optimization_level "medium" now uses 1 restart (was 2). Restart 1 is what finds the deeper basin on this dataset; restart 2 has never beaten it across the test runs. Final Lep timings: fast 3:17 (-63.4% score, no restarts) medium 4:16 (-65.9% score, 1 restart) slow ~10 min (5 restarts) thorough ~hour (8 restarts + snapshot PDFs) User-facing knob unchanged: optimization_level config key. Continues #127.
User-identified failure pattern on the Lep test: medium fixes a
visible block-swap between the top 7 rows that fast leaves behind --
the rightmost two chromosomes across rows 0..6 need to swap as a
unit. Single-row swap/insertion can't find this because moving just
one row's pair makes that row's crossings worse against the row
below; the synchronisation across all rows is what makes the move
a global win.
New move type: synchronised column-pair swap. For every adjacent
column pair (k, k+1) and every contiguous row span [r0, r1], swap
the chromosomes at columns k and k+1 in *every* row of the span
at the same time. Accept if global crossing score drops.
Runs as a fourth brushing phase after td + tm + bu + bm, snapshots
named ..._gen{N}_05_cs.pdf in thorough mode.
Cost: O(K * N^2) candidates per call (K = max columns, N = species
count) -- ~30 * 78 = 2300 candidates on Lep, each a single
total_crossings recompute. Adds ~5s per gen on Lep.
Effect: solves what restarts used to solve, in the very first
descent. Lep fast / medium results:
fast 3:17 / 389B -> 2:47 / 357B (-32B, ~15% faster)
medium 4:16 / 363B -> 3:45 / 354B (-9B, ~12% faster)
The new "fast" beats the old "medium" in both score and time -- the
col-swap move makes a single descent reach quality the old code
needed a perturbation restart to find.
Continues #127.
Fast mode used to leave a visible Cydia-Bombyx flip that medium caught via random restart -- the cascade's row-by-row greedy locks in a flip choice using an unweighted metric and the boundary pass can't single-flip-undo it because the fix requires multiple simultaneous flips. Three new final-polish passes catch this: 1. Iterated single-flip recheck on the global FET-weighted score. Cycles every (sp, scaf) up to 6 passes, accepting any flip that drops the total. Catches downhill cascades where each step is individually an improvement. 2. Pair-flip pass. For each FET-best (top_chrom, bot_chrom) pair, try flipping both simultaneously. Catches the case where neither flip alone improves the score but together they do. 3. Triple-flip pass. For each bot chrom B, find its top-2 FET partners P1, P2 in the row above; try flipping (B, P1, P2) together. Catches user-identified cascade patterns where one chrom in one row and two chroms in the row above need to flip in unison. Adds ~2 s to brushing wall time. Score impact on Lep fast mode: pre: 357B (-66.4%) post: 337B (-68.3%) i.e. better than the prior medium (354B), in the same ~3:42. Continues #127.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes #127.
True:Falsereproduces the previous plot byte-for-byte (verified on a 12-Lepidoptera test case: 889135 bytes identical).scripts/odp_rbh_to_ribbonused to carry an inline copy ofribbon_plot/_quality_check_chromosome_list/_optimize_spA_based_on_rbh/plot_bezier_linesthat had drifted out of sync withscripts/odp_ribbon_plot.py. The snakefile now imports the module directly.groupby(...).apply(...)+ multi-nameindex.namesassignment inribbon_plotbroke on pandas >=2 (ValueError: Length of new names must be 1, got 2). Replaced with an equivalentcumcount()-based construction of the within-scaffold gene index.plot_all: Truethealphacolumn was selected without being populated, raisingKeyError: "['alpha'] not in index". Now computed in both branches.Test plan
optimize_chrom_rotation: False, show_orientation_marks: False-> byte-identical to plot produced before this PR.optimize_chrom_rotation: True-> visibly fewer twists in the Lepidoptera plot; cascade completes in <2 min for 12 species, ~2k orthologs per pair.show_orientation_marks: True-> triangles render at the fasta-3' end of every chrom bar, on both the chr-coord and rbh-gene-coord panels.Notes / follow-ups (not in this PR)