[thebhef]

The owner did not accept the residual held-code loss: "It's ok to lose
codes that are extremely narrow, but if there was a transition there we
should make sure it shows up when zoomed way out." The plot serve
dropped a held code 1.49 columns wide on a level-1 read, and on a
level-3 read a run 2.2 columns wide (one served column is one canvas
pixel). Two causes, both in the serve, neither in the pyramid fold:

• decimate_min_max bucketed by point index. A pyramid level folds a
  flat bucket to one point and a busy one to two, so a bucket of points
  can be a fraction of a pixel in one place and several pixels in
  another, and a held run arriving as one point mid-bucket was neither
  first, last, min nor max. It now bins into max_points equal time
  columns over the window (clamped to the series' own span, so a
  fit-all serve over f64::MIN..MAX still works). It keeps each column's
  first, last, min and max, and the 4 * max_points bound is unchanged.
• A level's in-window count only bounds its average spacing. Beside a
  burst of changes a coarse level's gaps run to several columns, and a
  run with no point of the slice inside it lies inside one gap.
  SignalCache::refine_gaps splices finer-level points into every gap
  wider than a column, one level at a time, and stops as soon as no gap
  is wider. Every pyramid point is a real sample, so a run held for a
  column can no longer hide.

The serve now guarantees, in served columns: a run held for at least
one column is served with its code at every alignment and at every
level read. A narrower run may lose its code but not its change: the
serve changes value in that column or the one beside it. The one column
of placement is inherent to reading the pyramid, since a kept extreme
carries its own timestamp from anywhere in its bucket. Placing it
exactly needs a raw read at every column edge, which at a wide zoom is
the whole capture. On the frontend, drawEnumTiles draws a served run
narrower than a device pixel as a one-CSS-pixel tick in its accent,
the full band tall, on the lanes axis and the single-enum ribbon,
because a sub-pixel fill antialiases to nothing.

Measured on the real SignalCache (100 Hz cycling 0..=5, 30 130
samples). The fold and the pyramid (10 034 points above level 0) are
unchanged:

	largest held run lost	control serve	points read
before, 2248 pts (level 1)	1.49 col	5651	7536
after, 2248 pts	0.74 col	5986	7536
before, 280 pts (level 3)	>= 2.20 col	492	492
after, 280 pts	0.98 col	840	1896

Rejected after measuring:

• Reading a 4x finer level gets the change into exactly its own column
  in every sample tried, but costs +46 % serve and 4x the read.
• A per-point transition count in the fold costs 16 B per pyramid
  point plus a persistence change, still cannot place a change more
  precisely than its bucket, and does nothing for the held code.

ADR 0026 records the guarantee and the measurements. Tests:

• a_held_code_one_served_column_wide_survives_the_serve_at_every_alignment
  (two depths; red before);
• every_raw_change_shows_as_a_served_change_within_a_column (random
  holds and blips; it already held, now pinned);
• the sampler's column-edge sweep and a sparse-point case;
• the sub-pixel tick on ribbon and lane at ratio 1 and 2.

The perf bench's naive-path call to decimate_min_max takes the new
whole-range arguments.

zoom_in_reads_a_finer_level_than_fit_data asserted 2 * max_points + 2.
That bound has been stale since first/last joined the reduction (it
passed by luck of its sawtooth), so it now asserts the documented
4 * max_points.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com

[thebhef]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• Queue the feedback campaign's findings for owner review #502
• Startup restore feedback: a relaunch reports the prior capture's reopen in the status bar #575
• Plot y scale: fit what is on screen, keep a bound typed and left by clicking away #574
• Report a failed manual send and a failed Duplicate #573
• Round calendar-time fractions so they agree with the elapsed column #572
• Keep an enum's held codes and its changes visible at every zoom #571 👈 (View in Graphite)
• Adopt ruff 0.16's default rule selection in the three Python packages #570
• Duplicate a transmit element with its own cloned host TX frames #569
• Add a Duplicate button to the Project panel's element rows #568
• Read calendar time on the plot under time_display #567
• Read calendar time in the trace, By-ID, Events and Signals time cells #566
• Add the time_display setting and the calendar-time formatters #565
• Draw a ranged series straight between its own samples #564
• Copy plot areas with the platform's modifier, same panel included #563
• Keep a name beside a fixed sibling shrinking instead of vanishing #562
• Collapse plot axes only, with the disclosure leftmost on every head #561
• Logger {now} names each file of a run, not the run itself (task 165 phase 3) #560
• Date/time pattern setting: five display sites collapse into one formatter #559
• Name templates speak TR35 date patterns; strftime removed (ADR 0062) #558
• Remove the refused-send row: only the wire writes data (ADR 0061) #557
• Python toolchain: forward-and-consistent grpcio/protobuf/grpcio-tools, ruff 0.16 with an explicit rule set #556
• Row the time cursor chips when they collide; scale their precision with zoom #555
• Fault-recovery bench: six PEAK runs of the shipped recovery recorded #554
• Reopen a channel close-first, retry a failed open, and let a bus-off refusal arm the reset #553
• Add a fault-recovery bench that swaps the sidecar's bus-off recovery per run #552
• Persist bus-error episode detail with the error series #551
• Fix the host README's stale stuck-queue paragraph for ADR 0060 #550
• Render ADR 0060's bus-fault fields in the frontend #549
• Read bus faults from the sidecar's episode reports on the host #548
• Decode ADR 0060's bus-fault envelopes into per-interface client state #547
• Report bus faults from the sidecar as episodes, on a control lane ahead of data #546
• Add bus-fault reporting envelopes to cannet.v1 (ADR 0060 phase 3) #545
• Record the bus-fault model: sidecar-reported episodes, two lanes, flush a stuck queue #544
• Add a fault-load harness for the sidecar's session stream #543
• Remove the redundant Events disclosure chip from plot and trace #540
• List bus-error episodes among the events in the Events panel's one list #539
• Let only the wire write the Tx row: the driver's echo, not the send #538
• Refresh the plot's bus-error markers as errors arrive #537
• Add a cargo fmt --check job to CI #535
• Show searching… then the host's match count in the trace filter box #534
• Draw the plot's bus-error markers as episodes at the configured gap #533
• Read event and bus-error episode times as wall time on hover #532
• List bus errors as episodes at a configurable gap in the Events panel #531
• Page bus-error episodes into the Events panel; leave the trace's rows alone #530
• Give bus-error markers a windowed query on the plot #529
• Serve bus errors as a paged series from the signal cache #528
• Widen the logger file grid's action column and fix its column order #527
• Keep the window up while the host finishes closing #525
• Hold a project's cache directory exclusively for the session #524
• Reproduce the restore-time 1224 panic as two live mappings on one scratch #523
• Log a logger's clamped timestamps the way Save Capture already does #522
• Turn the trace store's before-session frame drops into a coalesced WARN #521
• Make PEAK bus-off visible and reopen a silent, queue-full channel #542
• Withhold PEAK echoes while error-passive; clear stale counters per poll #541
• Bring a bus-off controller back without user action, on every vendor #536
• Unwrap the Kvaser 32-bit receive timer in the sidecar driver #520
• Move the Servers rows into the settings view as a gridview #519
• Make the project caches list a gridview, and share its scroll-restore #518
• Make the settings view's units section a gridview #517
• Open the trace panel's signal disclosure on a signal or value winner #516
• Rank signals and enum values in their own right in the trace filter #515
• Answer before deriving: listings, index rebuilds and poll guards #514
• Take the heavy host commands off the IPC thread #513
• Alias a library unit through the settings units entry #512
• Colour a gridview selection at once, in Signals panel and plot area #511
• Project caches: chip carries location, tooltip follows it, no more danger button #510
• Filter the unit picker, and let a composed kind offer its whole class #509
• Read the whole unit table out of the library instead of typing it #508
• Name the project in each project-caches row; Delete becomes two-stage #507
• The settings view re-reads what it shows and keeps its place #506
• Draw plot sample markers as squares, collapsed per pixel column #505
• Draw the plot's hover chrome on an overlay canvas #504
• Close the palette before running the picked command #503
• Plumb empty plot areas, per-area event-marker base, and an Events chip #501
• Move the cursor and event readouts out of the data area into gutters #500
• Mark every served sample, and make an enum lane an ordinary series #499
• Serve every window as first, last, min and max, and drop the categorical reducer #498
• Measure what a plain min/max serve does to a held enum code #497
• Filter the Trace panel's rows with the same fzf, in both modes #496
• Rank the trace's rows host-side with the app's own fzf #495
• Let a bus explicitly set to no interface connect as unbound #494
• Let regular collapse/expand work while the database/RBS filter is active #493
• Groom five tasks: four out of the user feedback, plus the trace filter #492
• Fix fit-data test racing its own catch-up deadline #526
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.