Skip to content

processSEG 0.1.0: the processing library - #1

Merged
xrfez merged 61 commits into
mainfrom
feature/processing-library
Nov 11, 2026
Merged

processSEG 0.1.0: the processing library#1
xrfez merged 61 commits into
mainfrom
feature/processing-library

Conversation

@xrfez

@xrfez xrfez commented Nov 11, 2026

Copy link
Copy Markdown
Owner

The first release of processSEG: seismic processing in Nim on top of parseSEG (>= 0.8.0), built for an interactive viewer. Kernels work in place on parseSEG's sample tensors, chunk by chunk, with no hidden allocation and no hidden threads.

Contents

  • FFT: pure-Nim mixed-radix 2/3/4/5 real FFT, vectorized across traces.
  • Filters: zero-phase FFT band-pass (cos² trapezoid corners, Butterworth magnitude, notches) with edge handling to each trace's live length; Butterworth IIR, zero or minimum phase.
  • Gain: fixed dB, per-trace and per-panel normalization, AGC, time-variant curves, statistics, spectra.
  • Pipelines: ops compiled once to a plan, applied in place or src → dst, with the margin a zoom window needs; a persistent ProcessPool runs a plan across cores.
  • Pre-stack: mutes, datum statics, LMO and NMO, velocity functions and fields (VELF, HANDVEL, CSV), CMP binning, stacking with per-sample fold, a brute stack written as SEG-Y.
  • QC: geometry checks, first-break pickers with confidence, a per-shot refraction fit that flags swapped receivers.
  • Docs: README, guide pages with tested examples, API reference, integration contract.

Testing

35 test files pass locally in debug, release, danger and ASan builds; property and fuzz tests; benchmarks in bench/.

🤖 Generated with Claude Code

xrfez and others added 30 commits November 9, 2026 23:17
@xrfez @claude
0.5: Draft integration contract for parseSEG, processSEG and the viewer
dc97aec 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.5: Revise integration contract after design review
befdad9 
ChunkMeta gains live length, window origin, full geometry and signed
offsets; document the ORC/ptr threading rule, plan generations, edge
handling, SEG-D Rev 1/2 geometry via SPS, and a fuller P2 signature.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.6a: Propose parseSEG changes for a single shared trace data model
d3874b5 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
Move to parseSEG 0.7.0 (shared trace data model)
57794c7 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.6b: Record parseSEG 0.7.0 review follow-ups (F1-F3)
fe20ee9 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.6c: Patch parseSEG from here on the processseg-integration branch
11d012a 
processSEG.nimble requires the parseSEG branch processseg-integration;
nim parseSEGTest runs parseSEG's own suite in deps/parseSEG, and
nim parseSEGBranch re-attaches the checkout after atlas install.
proposedChanges.md is removed: C1-C8 shipped in parseSEG 0.7.0, and the
follow-ups are now processSEG tasks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.1: Rename the umbrella to src/processSEG.nim; recursive fmt
4d8b138 
- src/processseg.nim -> src/processSEG.nim (D9), build/test/docs tasks updated.
- `nim fmt` formats every .nim file under src/, tests/ and bench/.
- CI fmt check walks subdirectories (globstar) and reports every diff.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.2: Test runner discovering tests/t*.nim; nim asan task
b89ee32 
- tests/tester.nim compiles and runs each t*.nim in its own process, passing
  its arguments through as build flags (debug, -d:release, -d:danger,
  -d:addressSanitizer), and fails listing every failing file.
- tests/config.nims holds the shared paths and the ASan switches
  (-d:useMalloc, -d:noSignalHandler).
- Placeholder Context API removed; tumbrella.nim is the smoke test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.3: Synthetic signals and gathers (tests/utils/synth.nim)
b8e49d3 
- Signals added in place into var openArray[T]: sinusoid, spike, Ricker,
  DC, linear drift, seeded Gaussian noise.
- SynthGather[T]: parseSEG TraceKeys with geometry (signed offset, source/
  receiver/midpoint x, elevations, t0Ms, intervalUs, live status), per-row
  live lengths, a zero-padded [rows, cols] tensor, and a TraceRows view.
- Shot (split-spread, end-on) and CMP gathers; hyperbolic and linear events
  with optional per-trace shifts; elevation statics.
- tests/tsynth.nim self-tests every generator.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.4: Bench harness (nim bench) with a memory-bandwidth baseline
5a668a2 
- bench/benchutil.nim: warm-up + timed samples (each >= 20 ms), median and
  minimum per iteration, GB/s; prints CPU, OS, Nim, C compiler and the build
  flags with every run.
- bench/bmemory.nim: copy and in-place scale of a 2000 x 2001 float32 panel.
- `nim bench` runs bench/b*.nim in a danger build with recorded flags;
  results recorded in bench/README.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.6: Update the integration doc for parseSEG 0.7.0, the pool and one …
1af0c3d 
…data model

- ChunkMeta is a zero-copy wrapper over parseSEG's TraceRows (D12): per-row
  values from the view, liveness rule (tsLive/tsUnknown, not failed), and
  the three view lifetimes.
- Geometry from TraceKeys (inline/crossline from the layout), traceStatics,
  and keysRevision for anything derived from keys.
- Threads: the three ways to run, with the persistent public ProcessPool (D6);
  batch jobs in the prefetcher transform hook (D13); pure-Nim FFT (D2).
- "Proposed parseSEG changes" becomes "parseSEG 0.7.0 and the patch branch"
  (P1-P4 landed; F1-F3 on processseg-integration). Viewer checklist uses
  c.rows, uniform = true and keysRevision. Pool open question closed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.7: Pin parseSEG processseg-integration with TraceRows.keysAt (F1)
3ce4c9d 
parseSEG 6780373 adds keysAt(rows, row), the non-raising form of traceKeys
for {.raises: [].} kernels and ChunkTransforms (tested there; parseSEG's
whole suite passes through `nim parseSEGTest`).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
0.8: Phase 0 review gate fixes
20be299 
- integration doc: derived offsets (SEG-D Rev 3, SEG-2, survey joins) are
  unsigned distances, so split-spread side comes from the coordinates;
  kernels read keys with keysAt (F1 marked done); the adapter's interval
  check is a pass over the rows, not O(1).
- synth: unsignOffsets() models derived offsets; the rows() doc forbids
  temporaries.
- tester: runs the compiler with startProcess (no shell quoting).
- tests/config.nims: ASan flags for gcc/clang or MSVC; MinGW is an error.
- bench: the CPU target comes from config.nims (-d:cpuTarget) instead of
  being re-derived; clock read once per doubling batch; a compiler memory
  barrier after every body run (without it gcc merged batched in-place runs
  and reported 2x the real bandwidth).
- atlas.lock: parseSEG with the review's test fix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
1.1: ChunkMeta (core/types), windowed synthetic gathers, demo files f…
54a80c0 
…or tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
1.2: RowBlock (core/rows), Workspace (core/workspace), allocation cou…
518e940 
…nting and reject tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
1.3: Basic kernels (core/basic), mutes (core/mute), vectorization test
459035b 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
1.4: Adapters (chunkMeta from chunks and reads); pin parseSEG F2/F4
098b707 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
1.5: Phase 1 review gate fixes
2307f28 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
2.1: Real mixed-radix FFT (fft/plan), batched across traces; bench
cb47af4 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
2.2: Amplitude spectra and dB for display (fft/spectrum)
34ff29c 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
2.3: Phase 2 review gate fixes
8fe69b4 
NaN/Inf rows left out of averaged spectra; always-on FFT buffer length
checks; short windows tapered and scaled per row; default-plan assertion;
workspaceSize(batched); more lane, NaN and float64 tests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
2.4: Power-of-two spectrum nfft by default; 60 s at 1 ms records test…
16efac6 
…ed and benched

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
3.1: Filter responses (filter/design): cos² trapezoid, Butterworth, n…
25bf53e 
…otch, impulse length

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
3.2: Zero-phase FFT filtering with D4 edge handling (filter/apply)
47407ef 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
3.3: Butterworth biquad cascades (filter/iir), zero and minimum phase…
97b3dae 
…; warped FD Butterworth

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
3.4: Filter bench (2000 × 2001 band-pass, 1 and 4 workers); viewer bu…
92a9ba7 
…dget

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
3.5: Phase 3 review gate fixes (per-feature impulse length, min-phase…
355a693 
… margin, zeroDc)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
4.3: Trace and panel amplitude statistics (gain/stats), percentile cl…
193aab8 
…ip by histogram

TraceStats/traceStatsInto, PanelStats/panelStats (65536-bin float32-bit histogram,
0.4 % relative), exact percentileAbs; Workspace gains a uint32 region. Bench in
bench/bgain.nim. Also applies nim fmt to the Phase 3 filter files (layout only).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
4.1: Fixed dB gain, normalization and AGC (gain/gain)
99bdfe1 
scaleDb; NormalizePlan (normMax/normRms/normPercentile, perTrace/perPanel);
AgcPlan (agcRms/agcMeanAbs, centred window, real counts at the ends,
epsilon relative to the trace RMS, live length, kept gains + removeGain).
Window sums from block suffix + prefix sums: no subtraction, exact to a few
ulps at any dynamic range, O(n) whatever the window. tvectorized groups a
kernel's GCC parts and clones.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
4.2: Time-variant gain curves (gain/curves)
fc78726 
GainCurve (knotGain linear in dB and flat outside, powerGain, exponentialGain),
GainCurvePlan tabulated on the dt grid; applyGainCurve multiplies each live row
by the table at its own times (t0Ms, firstSample), linear between grid points
for a fractional shift, the curve itself outside the table.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
xrfez and others added 29 commits November 10, 2026 20:43
@xrfez @claude
5.2: ProcessPool (persistent workers, one Workspace each) and nim asa…
612bad9 
…n in CI

Workers start once, sleep on a condition variable and are joined by
=destroy; =copy and =dup are errors. apply validates on the caller's thread,
splits the rows into one range per worker, and runs a per-panel factor on
worker 0 between parts. Identical to the single-thread apply for 1-8
workers; ASan and TSan clean. Bench: 6-12 us per job, 7.4x at 8 workers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
5.3: Constant-velocity moveout times for overlay ticks (velocity/move…
40efc5b 
…out)

lmoTime/nmoTime, Moveout (linearMoveout, hyperbolicMoveout), and
moveoutTimesInto (ms from a TraceRows view) / moveoutColumnsInto (block
columns from a ChunkMeta). Offsets keep their sign in the keys; the formulas
use |x|. NaN offsets or t0 give NaN.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
5.5: Viewer simulator (test code only) and the integration doc on the…
42197af 
… real API

tests/utils/viewersim.nim acts out a viewer: uniform prefetch into a raw
cache, ProcessPool into processed buffers tagged by generation, a checking
draw, slider edits, zoom windows with requiredMargin, moveout ticks tagged
with keysRevision and made stale by applySurvey. tviewerloop drives it on
synthetic data and parseSEG's demo files: stale buffers always reprocessed,
0 allocations per steady-state frame, ASan clean.

The doc now names the real API (ops, Pipeline, apply, pool, moveout) and
says what the margin promises: relative to the strongest nearby signal; AGC
after a filter lifts the leak in quiet windows; whole-trace ops use the
window's statistics.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
5.6: Phase 5 review gate fixes (key staleness, API friction, pool and…
3060583 
… pipeline checks)

Two fresh reviews: a docs-only API review that wrote its own viewer loop
(kept as tests/tapireview.nim) and a diff review.

- Processed buffers are stale when the plan's generation or the file's
  keysRevision changes (mutes read offsets, gain curves t0, every op the
  status); viewersim tags both and gains refresh(f); new viewer-loop test.
- pipeline: factors checked against the sample type; checkApply(src, dst)
  rejects partial overlaps; runPart/panelFactorOf assert their preconditions
  (kept public for viewers with their own threads); muteTopOp/muteBottomOp;
  requiredMargin(flow, dtMs) without compiling; decodeWindow.
- pool: workers of a moved-from pool is 0, ping asserts, a failed
  createThread joins only started threads.
- moveout: hyperbolic t0Ms >= 0; moveoutTimesInto(meta, ...).
- The umbrella re-exports parseSEG. Integration doc: cached-chunk lifetimes,
  what allocates, zoom windows on the traces' peak, batch marked planned.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
Keep parseSEG out of the processSEG umbrella
e021c89 
import processSEG no longer re-exports parseSEG: the two packages stay
separate, and callers import both (import parseSEG, processSEG). parseSEG
is on every build's path through Atlas's nim.cfg.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
5.2b: tlanif model of the pool's job hand-off and shutdown
4668a1a 
models/pool_handoff.nif: two workers, up to two jobs, destroy at any idle
point, spurious wake-ups. Every job runs exactly once per worker before the
caller returns, and no wake-up is lost. Base passes (235 states), the
witness is reachable, and all five mutants fail (no recheck after a wake,
wait ignoring stop, caller waiting once, no done signal, signal instead of
broadcast).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
6.1: Velocity functions, Dix conversion, 2D/3D fields and VELF/HANDVE…
65c8945 
…L/CSV picks

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
6.2: Moveout from velocity models, LMO apply/remove and fractional sh…
22a9c98 
…ifts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
6.3: Phase 6 review gate fixes (grid overflow, VELF/HANDVEL/CSV mispa…
3989441 
…rses, zero knot, negative gradient)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.1: Datum statics from elevations and header columns, applied as fra…
e7c654a 
…ctional shifts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.2: NMO correction from each row's t0, linear/cubic, with a tapered …
dcd2d17 
…stretch mute

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.3: CMP binning: header CDP, line projection, rotated 3D grid, heade…
c1b4f2c 
…r inline/crossline

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.4: StackAccumulator: any order, per-sample fold, fold/sqrt-fold nor…
617348c 
…malization, merge

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.5: bruteStack over a file in the prefetch transform: statics, flow,…
e628ab6 
… NMO, per-worker stacks, progress and cancel

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.5: nmo doc substitution warning
17c84c6 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.6: Phase 7 review gate fixes (decode failures counted, NaN statics …
ea09ee7 
…and intervals skipped, live-sized stack, worker early stop, overflow guard, rcvDepth, header statics column)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
7.6a: writeStack: a stack of any input standard as SEG-Y (parseSEG F3…
a03bfd7 
…: traces from keys)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
8.1: geometry QC: header offset vs coordinates, zero/negative/missing…
9adac1f 
… offsets, midpoint, per-trace flags

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
8.2: first-break picking: STA/LTA, modified energy ratio, modified Co…
4e2bb47 
…ppens; minimum-phase band-pass, LMO-guided window, SNR confidence

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
8.3: per-shot refraction fit (two-segment L1 or offset-binned), local…
c53443e 
… residuals, neighbour cross-correlation, symmetry; swapped receivers flagged

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
8.4: Phase 8 review gate fixes (spread order without coordinates, NaN…
f2eb5fa 
… t0 picks, degenerate and long-spread fits, midpoint flag out of badGeometry, writeStack sorting and docs, detection limits)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.1: property and fuzz tests; fix float-to-int conversions of out-of-…
8033777 
…range values

tests/tproperty.nim: random pipelines on random gathers (sizes, dt, live
lengths, dead/auxiliary/failed rows, 1e-6..1e6 amplitudes): compile raises
only ValueError, apply is finite, leaves non-live samples alone and equals
src -> dst; a NaN/Inf sample reaches no other row; FFT and IIR filter
linearity; AGC and normalize scale invariance; zero in, zero out; shifts,
NMO and first-break picks with extreme parameters; velocity parsers on
mutated text.

tests/fuzz: libFuzzer + ASan + UBSan targets for the velocity parsers and
for pipelines built from bytes (raw float parameters and samples), with
seeds and a `nim fuzz` task.

UBSan findings fixed: the impulse-length grid of a near-zero ramp (an
18 s run on a garbage FFT size), the NMO taper start for a huge taperMs,
and first-break and shot-fit correlation windows of huge finite lengths.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.2: performance pass on the pre-stack and QC kernels (bprestack benc…
f4d2eee 
…hmark)

- first breaks: a band-passed plan filters fftLanes rows at a time through
  the IIR lanes kernel into workspace copies (was one scalar trace at a
  time): 5-60 Hz picking 100/86/91 -> 44/31/35 ms per 2000-trace panel;
  unfiltered plans pick from the rows in place;
- NMO: the RMS velocity column is reused across rows with the same start
  time (and location), and the input columns are computed by a vectorized
  block loop (moveoutCols, -fno-math-errno in the nmo module): linear
  36 -> 17.5 ms, cubic 55 -> 25 ms;
- shot fit: l1Fit normal equations in local sums (vectorized), golden
  section reuses one point per step: 38 -> 21 ms on 2000 traces;
- tvectorized checks moveoutCols and l1Fit; bench/README records the run.

nim test and nim testRelease pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.3: docs: README, guide pages with tested examples, API reference
c98a03c 
- README: overview, a compiled quick start, links;
- docs/guide/: getting started, pipelines (zooms, the pool, own threads),
  gain, velocity and stack, QC; every example is a program;
- docs/api.md: conventions, then every exported name by module;
- tests/tdocs.nim: compiles and runs the guide and README examples in
  one build, and checks api.md names all 370 exported names (jsondoc);
- docs/integration.md: status for the release candidate, QC workspace,
  and the zoom margin's measured worst case (2.5 %, not 1 %);
- doc comments: nim docs is warning-free.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.4: whole-library review fixes (three fresh reviewers by area)
c0163e5 
Defects:
- mute kernels: opposite infinite bounds made a NaN ramp bound and the
  boundChecks-off loops ran from int.low (heap overflow under ASan); an
  infinite taper is now no taper, the column helpers are NaN-safe, and a
  subnormal dtMs is refused by initChunkMeta;
- intervalVelocity with a first time of 0 changed the first layer (a knot
  at 0 makes Vrms linear from it): the zero-thickness layer is dropped;
- float32 overflow to Inf turned zeros into NaN: gain-curve tables and
  scaleDb refuse such gains, block scale refuses a factor Inf in T, and
  off-table curve samples keep zeros zero;
- FFT/IIR filter and averaged-spectrum block forms now refuse rows whose
  interval is not the plan's (they filtered at the wrong frequencies);
- first breaks: an infinite t0 is no time axis (was a NaN-to-int);
- shot fit: golden-section refinement kept within the candidates' range;
  fitShot no longer infers raises: [Exception] (forward declaration);
- rowBlock(data, rows, cols) refuses an overflowing size;
  decodeWindow takes a negative margin as 0 and cannot overflow;
- velocity parsers accept a UTF-8 byte-order mark.

API: pick(plan, x, ws, offset = NaN, t0Ms = 0) (ws no longer after the
defaults). Docs: zero-mean claim of withZeroDc, asymmetryMs sign, lag
choice, MER weight, pool exceptions, perPanel in bruteStack is per chunk,
the zero-offset tolerance; doc comments from the docs pass. Tests for each
fix, an NMO reference test, allocation-free filtered pickInto.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.4: pin parseSEG e3b855d (review fixes to the writer from keys)
ab4e841 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.5a: require parseSEG >= 0.8.0 (released from the patch branch, PR #1)
9c0f2c2 
The patch branch is merged upstream and released as parseSEG 0.8.0
(v0.8.0, 7be5c47); atlas.lock pins it, so CI's atlas rep resolves a
pushed commit again. Integration doc: parseSEG section for 0.8.0, and
the zoom margin's worst case as an open question for the viewer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.5: CI formatting: nimpretty at 100 columns, sources formatted
7ab7642 
nimpretty's default 80 columns rewrapped the 100-column sources and, on one
wrap, joined `and p.lagToleranceMs` into `andp.lagToleranceMs`. CI and
`nim fmt` now pass --maxLineLen:100; the remaining diffs are comment
alignment. tfilterapply's spec list is hoisted out of the for header (a
comment inside it crashed nimpretty), and two long test lines are rewrapped
by hand.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez @claude
9.5: scale two float tolerances to the data (gcc 15 release CI)
e6320d4 
On the CI runner (gcc 15, x86-64-v3) the release build differs from gcc 12
in the last bits: tpipeline's plan-vs-ops check asked for bit equality
(off by 6.7e-16), and tproperty's NaN-containment check compared a
near-zero sample (2e-11) relative to itself. Both tolerances are now
relative to the data's scale (the output's maximum, the row's maximum).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@xrfez
xrfez merged commit 6903084 into main Nov 11, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@xrfez