edfcore 0.2.46 → 0.2.48

package/CHANGELOG.md

@@ -6,7 +6,37 @@ alone does not tell you whether you were affected.
 edfcore is pre-1.0. Patch releases have carried behaviour changes where the old behaviour was a
 defect; those are called out below.
 
-## 0.2.45
+## 0.2.48
+
+- **Added** bit-for-bit parity against pyEDFlib on the REAL corpus — the last thing the README
+  withheld. A 22-hour clinical polysomnogram from sleep-edfx and the three teuniz generator files
+  in EDF, EDF+ and 24-bit BDF+, compared with `Object.is` per sample.
+- These files matter because nobody here chose them: they were written by other people's software
+  and hardware, years ago. `corpus.test.ts` already read them, but it checked that the output was
+  BELIEVABLE — a rectal temperature near 37 degrees, an 8.5 Hz channel oscillating at 8.5 Hz — and
+  a check like that would pass for a reader that was slightly wrong everywhere, which is exactly
+  the failure the pinned scaling expression exists to prevent. A test in the new file demonstrates
+  that directly: on the PSG's temperature channel the textbook expression produces a believable
+  body temperature for every sample AND disagrees with pyEDFlib, so the old check could not have
+  told the two apart and the new one can.
+- Sampled at the start, the middle and the END of each signal rather than whole files — the PSG
+  alone is 48 MB. The end window is the one that earns its place: a reader whose record arithmetic
+  drifts does so with distance from the start.
+- Skips without the corpus, like every other test in that directory, so a fresh clone stays green
+  and offline.
+
+## 0.2.47
+
+- **Fixed** this changelog. Every heading from `0.2.36` down to `0.2.45` named a version one lower
+  than the release that actually shipped it: `0.2.36` was consumed by a release run that failed its
+  typecheck after bumping the version, and I kept writing the next entry against the number I
+  expected rather than the tag that got cut. So the `cachedSource` fix was labelled `0.2.36` and
+  shipped in `0.2.37`, and everything after it was off by one — including the MNE claim in the
+  README, which named `0.2.43` for work that is in `0.2.44`. Verified against the tags rather than
+  reasoned about: `git show <tag>:CHANGELOG.md` for each. `0.2.36` is now listed as never released,
+  the way `0.2.29` already was.
+
+## 0.2.46
 
 - **Added** `scripts/golden/README.md`: how to regenerate every parity fixture, and — more usefully
   — what each of the three harnesses actually claims and how strong that claim is. Bit-for-bit for
@@ -18,7 +48,7 @@ defect; those are called out below.
 - **Documented** the same table on the physical-values page, so a reader who never opens the
   repository sees which claims are load-bearing and which are approximate.
 
-## 0.2.44
+## 0.2.45
 
 - **Added** annotation parity against pyEDFlib — the other axis, and the one edfcore has got wrong
   most. The scaling harness checks arithmetic; this checks WHICH AXIS an onset is on. Six releases
@@ -33,7 +63,7 @@ defect; those are called out below.
   recorded from that would have made the parity test compare an incomplete set and pass while doing
   it.
 
-## 0.2.43
+## 0.2.44
 
 - **Added** parity against MNE — a second, independent reader. pyEDFlib and edfcore both descend
   from EDFlib's arithmetic, so agreement between them shows edfcore copied it correctly rather than
@@ -49,7 +79,7 @@ defect; those are called out below.
   other. Channels MNE does not rescale — it leaves `degC` alone — are excluded rather than pushed
   through a factor that would make the comparison an artefact of the test.
 
-## 0.2.42
+## 0.2.43
 
 - **Added** three cases to the pyEDFlib parity harness, chosen for where the two scaling
   expressions diverge most or where a mistake would be least visible:
@@ -61,7 +91,7 @@ defect; those are called out below.
     set.
   - **a full 16-bit range** mapped to one millivolt — the finest.
 
-## 0.2.41
+## 0.2.42
 
 - **Added** a `samplesPerRecord` column to `edfcore signals`, and **fixed** the documentation,
   which claimed the command emitted it when it emitted `kind` instead. That left the authoritative
@@ -72,7 +102,7 @@ defect; those are called out below.
   five columns by position moved. The full six-column list is now pinned by a test, not only
   described.
 
-## 0.2.40
+## 0.2.41
 
 - **Added** the types `edfcore/validate`'s own signatures mention to that subpath's exports:
   `EdfHeader`, `EdfDiagnostic`, `EdfDiagnosticCode`, `EdfSeverity`, `EdfSignal`, `EdfRecordIndex`
@@ -83,7 +113,7 @@ defect; those are called out below.
   of anything they passed or got back. A type-only test now pins the set, so it fails at
   `npm run typecheck` rather than in a consumer's project.
 
-## 0.2.39
+## 0.2.40
 
 - **Added** tests pinning which date defect produces which diagnostic. `DATE_IMPLAUSIBLE` is
   documented as covering two conditions with only the second reachable, and that was prose about an
@@ -94,7 +124,7 @@ defect; those are called out below.
   dead. The branch itself is kept, with the reason written next to it — a missing guard is harder
   to notice than an idle one.
 
-## 0.2.38
+## 0.2.39
 
 - **Fixed** `onsetSecondsFromFirstRecord` and `onsetTicksFromFirstRecord` disagreeing about the
   same event at the edge of the int64 tick range. The exact field saturated and the float field was
@@ -103,7 +133,7 @@ defect; those are called out below.
   with the exact field in 0.2.10. Both are now derived from one rebased value, which is what the
   float field was always documented to be: the lossy view of the exact one.
 
-## 0.2.37
+## 0.2.38
 
 - **Fixed** `fileHandleSource` and `fileSource` ignoring an abort signal that flipped while a read
   was in flight. The abort check ran at the top of each loop iteration, and the common case is one
@@ -112,7 +142,7 @@ defect; those are called out below.
   test saying why; one adapter honouring a signal that another quietly ignores is worse than either
   rule alone.
 
-## 0.2.36
+## 0.2.37
 
 - **Fixed** `cachedSource` repopulating itself after `close()`. A read already in flight when close
   was called still resolves, and its continuation still runs — after `blocks.clear()` — so the
@@ -126,6 +156,12 @@ defect; those are called out below.
   cache entirely and retains nothing. The path is exactly as safe as calling the wrapped source
   directly, which is what it does.
 
+## 0.2.36
+
+Never released. A release run failed its typecheck after bumping the version, which consumed the
+number before a tag was cut — the same way `0.2.29` was lost earlier. The `cachedSource` fix that
+carried this heading while it was being written shipped in `0.2.37`, and is listed there.
+
 ## 0.2.35
 
 - **Fixed** `byteSource` building a source over an argument that is not bytes, so the caller's

package/README.md

@@ -299,15 +299,25 @@ back with pyEDFlib, and records every sample as its exact IEEE-754 bit pattern;
 `tests/corpus/golden/` was produced by edfcore. Substituting the numerically better textbook
 expression fails it on 140 of 256 samples — which is why the EDFlib form is pinned.
 
-**MNE, as of 0.2.43.** edfcore agrees with MNE to within 1e-12 relative — a weaker claim than the
+**MNE, as of 0.2.44.** edfcore agrees with MNE to within 1e-12 relative — a weaker claim than the
 one above, and deliberately so. MNE returns SI units, so a microvolt channel arrives divided by
 1e6, and that division is lossy; the residue is MNE's unit conversion rather than a disagreement
 about the sample. Bit-parity is claimed for pyEDFlib alone. Both readers also agree with the file's
 own declaration about POLARITY on a negative-gain channel, which a value comparison alone could not
 establish.
 
-**Still open before 1.0.** Validation across public corpora (sleep-edfx, CHB-MIT, the BioSemi and
-edfplus.info test files). That claim is not made here until a test produces it.
+**The public corpora, as of 0.2.48.** The bit-for-bit claim above also holds on files nobody here
+chose: a **22-hour clinical polysomnogram** from sleep-edfx, and the teuniz generator files in EDF,
+EDF+ and 24-bit BDF+. Sampled at the start, the middle and the end of every signal — the end window
+being the one that catches record arithmetic drifting with distance, which a sample near record 0
+cannot. Those tests need the corpus and skip without it:
+
+```bash
+npm run corpus:fetch
+```
+
+**Still open before 1.0.** CHB-MIT, and the edfplus.info conformance files. Those are not claimed
+here until a test produces them.
 
 **Shipped since 0.1.6.** Min/max envelope decimation (`readEnvelope`,
 `readEnvelopeAtResolution`). BioSemi Status-byte helpers (`readTriggers`). Streaming iteration

package/dist/cli-run.js

@@ -221,7 +221,7 @@ export async function runCli(args, io) {
         case 'signals': {
             // Tab-separated and one line per signal: `header` is for reading, this is for piping.
             //
-            // `samplesPerRecord` is last and was missing until 0.2.41. It is the AUTHORITATIVE field —
+            // `samplesPerRecord` is last and was missing until 0.2.42. It is the AUTHORITATIVE field —
             // `sampleRateHz` is derived from it and the record duration, and is empty for the legal
             // zero-duration file — so a listing meant for a script that omitted it forced the reader
             // back to `json` for the one number they could safely index by. Appended rather than

package/dist/constants.d.ts

@@ -109,5 +109,5 @@ export declare const SIGNAL_FIELD_BLOCK_OFFSETS: {
     readonly reserved: 224;
 };
 /** Published package version. Kept in sync with package.json by a test. */
-export declare const VERSION = "0.2.46";
+export declare const VERSION = "0.2.48";
 //# sourceMappingURL=constants.d.ts.map

package/dist/constants.js

@@ -79,5 +79,5 @@ export const SIGNAL_FIELD_BLOCK_OFFSETS = {
     reserved: 224,
 };
 /** Published package version. Kept in sync with package.json by a test. */
-export const VERSION = '0.2.46';
+export const VERSION = '0.2.48';
 //# sourceMappingURL=constants.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "edfcore",
-  "version": "0.2.46",
+  "version": "0.2.48",
   "description": "Modern, typed, zero-dependency reader for EDF, EDF+, BDF and BDF+ biosignal files. Works in browsers and Node with true random access.",
   "keywords": [
     "edf",

package/src/cli-run.ts

@@ -263,7 +263,7 @@ export async function runCli(args: Args, io: CliIo): Promise<number> {
     case 'signals': {
       // Tab-separated and one line per signal: `header` is for reading, this is for piping.
       //
-      // `samplesPerRecord` is last and was missing until 0.2.41. It is the AUTHORITATIVE field —
+      // `samplesPerRecord` is last and was missing until 0.2.42. It is the AUTHORITATIVE field —
       // `sampleRateHz` is derived from it and the record duration, and is empty for the legal
       // zero-duration file — so a listing meant for a script that omitted it forced the reader
       // back to `json` for the one number they could safely index by. Appended rather than

package/src/constants.ts

@@ -93,4 +93,4 @@ export const SIGNAL_FIELD_BLOCK_OFFSETS = {
 } as const;
 
 /** Published package version. Kept in sync with package.json by a test. */
-export const VERSION = '0.2.46';
+export const VERSION = '0.2.48';