Release 0.4.1: NoisePy co-install + last obspy call off the CCF path

CHANGELOG.md

@@ -4,7 +4,68 @@ All notable changes to seisfetch are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
 [Semantic Versioning](https://semver.org).
 
-## Unreleased
+## 0.4.1 — 2026-08-17
+
+The release NoisePy's obspy-free data path depends on: it makes seisfetch
+installable alongside noisepy's pinned S3 stack, and closes the last two gaps
+between the numpy pre-processing chain and noisepy's obspy one.
+
+### Added
+
+- `contrib.obspy_ports.gps2dist_azimuth_np`: port of
+  `obspy.geodetics.base.gps2dist_azimuth` (the Vincenty-inverse branch, the
+  one that runs in a default install). noisepy's `cc_parameters` writes
+  dist/azi/baz into every saved cross-correlation, so this was the last
+  obspy call on the cross-correlation path. Exactly equal to obspy's own
+  Vincenty on 2000 random coordinate pairs plus station-like, equatorial,
+  identical-point and longitude-wrapping cases; see the geographiclib note
+  under Fixed for the one install where "exactly" becomes "to ~5e-6 m".
+
+### Fixed
+
+- `contrib.noisepy_adapter.preprocess_raw_np` now rounds the merged start
+  time to whole microseconds, as obspy's miniSEED reader does, so noisepy's
+  sub-sample correction (derived from `UTCDateTime.microsecond`) cannot see a
+  different value than the obspy chain would. The rounding uses integer
+  arithmetic: at epoch-nanosecond magnitudes a float quotient carries a
+  0.25 ulp, which quantizes the sub-microsecond part to 250 ns steps and
+  picks a different microsecond for ~13 % of arbitrary inputs, and `round()`
+  is half-to-even besides.
+
+  Honest scope: this is a guard, not an observed repair. Scanning 3026
+  segments across all cached SCEDC, NCEDC and EarthScope day files found
+  **zero** with a sub-microsecond start — pymseed 0.9.4 reports whole
+  microseconds on every one, so the line is a no-op on all data available
+  here and both roundings agree with obspy on every fixture. It costs
+  nothing and closes the hole if sub-microsecond starts ever appear.
+- The `gps2dist_azimuth_np` equivalence tests asserted exact equality against
+  whichever branch obspy happened to take, so they failed on any machine with
+  `geographiclib` installed — obspy delegates to it when present and only
+  falls back to its own Vincenty otherwise. The exact-equality tests now skip
+  when `HAS_GEOGRAPHICLIB`, and a tolerance test runs either way so coverage
+  is not silently lost. The two branches differ by ~5e-6 m over ~500 km, and
+  geographiclib spells a due-south back-azimuth `0.0` where Vincenty says
+  `360.0`; both are far below what CCF metadata resolves, but it means
+  bit-identity holds against a default obspy install rather than every one.
+- `preprocess_raw_np(..., pretrimmed=False)` skips the sample-grid alignment
+  guard. The guard exists for callers that pre-trim segments with
+  `TraceBundle.trim` (inside-window) before pre-processing; a caller that
+  passes whole segments — exactly what `obspy.read` hands noisepy — is
+  trimmed only by `trim_pad0_np` at the end of the chain, which IS obspy's
+  nearest-sample trim, so unaligned windows are handled identically. NoisePy
+  asserts bit-identity on a window 0.218 samples off the grid.
+
+### Changed
+
+- `boto3` lower bound relaxed to `>=1.26` (was `>=1.28`). NoisePy cannot
+  install seisfetch otherwise: `noisepy-seis-io` pins `s3fs==2023.4.0`,
+  which forces `aiobotocore 2.5.2` and `botocore<1.29.162`, a range
+  `boto3>=1.28` excludes. `seisfetch/s3.py` uses only `boto3.client`,
+  `boto3.Session`, list-objects paginators, `botocore.UNSIGNED` and
+  `ClientError` — all long-stable. Verified on boto3 1.26.161: live
+  SCEDC/NCEDC/EarthScope day-file pulls and the full precision suite
+  (56 passed, including both NoisePy CCF equivalence harnesses,
+  bit-identical).
 
 ## 0.4.0 — 2026-08-16
 

docs/noisepy-obspy-replacement-report.md

@@ -88,6 +88,43 @@ on the S3 path, and `num_segments` counted records. Fixes:
   guard; lazy transport imports (`import seisfetch` no longer pulls boto3);
 - CI (there was none), five small committed fixtures, 20+ new tests.
 
+## Status (2026-08-16): the migration this report recommended is implemented
+
+All three steps below are done, on branches: `seisfetch` 0.4.1,
+`noisepy-io` `feat/seisfetch-input-layer`, `NoisePy` `feat/seisfetch-data-path`.
+
+| Step | State | Evidence |
+|---|---|---|
+| (i) adopt the numpy ports behind the `rm_resp` switch | done — `noise_module.preprocess_raw_np`, dispatched in `correlate.preprocess`; the ports are IMPORTED from seisfetch, not copied, so NoisePy (MIT) stays free of the LGPL file | `NoisePy tests/test_preprocess_equivalence.py`: bit-identical to the obspy chain on both fixture sets x 40/20/1 Hz |
+| (ii) array-backed `ChannelData` with a lazy `stream` | done — new `TraceSegment`; `ChannelData.from_segments`/`from_array`; `stream` built on first access | `noisepy-io tests/test_seisfetchstore.py` |
+| (iii) demote obspy to an extra + split the input layer | done — every obspy/pyasdf import in noisepy-io and NoisePy is now lazy; new `noisepy.seis.io.seisfetchstore` (routing owned by `seisfetch.s3`, coordinates from the FDSN text service); `obspy`/`asdf` extras in both packages | `NoisePy tests/test_no_obspy.py` runs read -> preprocess -> FFT -> cross-correlate in a subprocess where importing obspy RAISES |
+
+Two gaps this report did not anticipate, both found and closed during the
+migration:
+
+- **`cc_parameters` called `obspy.geodetics.base.gps2dist_azimuth`** for every
+  station pair, writing dist/azi/baz into every saved CCF. Ported to
+  `contrib.obspy_ports.gps2dist_azimuth_np` (Vincenty inverse), exactly equal
+  to obspy on 2000 random pairs. `noise_module.taper` likewise called obspy's
+  hann entry point; it resolves to `scipy.signal.windows.hann`, now called
+  directly.
+- **Microsecond rounding.** obspy's miniSEED reader rounds sample times to whole
+  microseconds and noisepy reads `fric` off `UTCDateTime.microsecond`; pymseed
+  keeps exact nanoseconds. On a day file starting at 00:00:00.019537920 the two
+  chains differed by ~1e-8 relative — not bit-identical. `preprocess_raw_np`
+  now rounds to microseconds, and the sample-grid alignment guard applies only
+  to pre-trimmed input (`pretrimmed=False` for whole segments, which is what
+  the obspy path gets from `obspy.read`).
+
+The **two-env caveat is now retired**: `tests/test_no_obspy.py` blocks obspy at
+import time and still completes a cross-correlation, so the claim is no longer
+"the data path does not use obspy" but "the pipeline runs without it installed".
+
+One dependency consequence: `noisepy-seis-io` pins `s3fs==2023.4.0`, which
+forces `botocore<1.29.162`, so seisfetch's `boto3>=1.28` floor made the two
+uninstallable together. Relaxed to `>=1.26` in 0.4.1 (verified on boto3
+1.26.161 against live SCEDC/NCEDC/EarthScope).
+
 ## Architecture recommendation
 
 1. **seisfetch becomes the sole owner of data-center knowledge.** Today