Skip to content

Releases

XinRefinement v1.7.0 — Amorphous hump component + recipe-path fixes

Choose a tag to compare

@XinProf XinProf released this 14 Nov 07:04

XinRefinement v1.7.0 — Release Notes

Released: 2026-11-14

Highlights

Amorphous hump component (opt-in)

Samples that contain a genuinely amorphous fraction (glass, fumed silica)
scatter a broad hump that a Bragg-only model cannot carry. The common
workaround — a denser background — silently eats intensity from overlapping
peaks. v1.7.0 adds a physical broad component instead:

  • RefineConfig.amorphous = ((center_deg, fwhm_deg), ...) and amorphous_eta
    (pseudo-Voigt mixing, default 0.5); CLI --amorphous 22:9 (repeatable /
    comma-separated) and --amorphous-eta; the GUI background dialog gains an
    "amorphous hump" group.
  • The hump amplitude is solved in the linear layer together with phase
    scales and background, so it adds zero nonlinear parameters and cannot
    steal degrees of freedom from the peak model. Center and FWHM are
    user-supplied and intentionally not refined.
  • Disabled by default; with amorphous=() every code path is bit-identical
    to v1.6.4 (golden CLI Rwp still 3.4074156640940005).

Recipe-path fixes

  • refine_asym is now a real switch (conclusion 7). RefineStep.config_for()
    used to hard-code refine_asym=(key == "profile") and ignore
    RefineConfig.refine_asym, so passing refine_asym=False to RecipeRunner
    still refined the asymmetry term (Rwp identical to the flag being on). The
    field is now tri-state: None (default — reproduces the old numerical
    path bit-for-bit), True (force on), False (force off). CLI gains
    --no-refine-asym.
  • K30 fixed — refine_eta was a dead parameter on the anisotropy path.
    When Popa LB anisotropy is active, forward._accumulate_peaks() took η from
    tch_eta() and ignored ProfileParams.eta, yet the profile recipe step
    still refined it (sweeping η 0→1 changed Rwp by exactly 0 pt; all 7
    benchmark samples run 100% through that TCH branch, and the GUI ships with
    aniso_auto enabled, so the dead parameter sat on the product default
    path). nonlinear.eta_is_active() now probes the cached peak list at
    refine() entry and clears refine_eta when no scalar-branch peak exists.
    The computed pattern, Rwp and every refined parameter are bit-identical on
    all 7 samples
    ; only n_params drops by one, which corrects Rexp
    (previously low by 0.0007–0.0015).

Honest benchmark notes (sample 7)

All of the following experiments were negative and are documented for
transparency rather than shipped as wins:

  • Adding the amorphous hump at its known 22°/9° position: +0.109 pt Rwp
    with QPA drifting 2.7 vol% → not adopted as a benchmark improvement.
  • A two-component "sharp core + broad wing" profile adds no expressive power
    beyond a single pseudo-Voigt using only existing knobs (global FWHM ×0.85
    plus η +0.3 reaches −1.379 pt). Rejected.
  • ka_ratio scans top out at −0.142 pt (at 0.6); locking QPA retains 98% of
    the dual-shape gain — so neither the Kα doublet nor phase-scale freedom is
    the bottleneck.
  • Fixed-QPA + peak-shape check: free QPA 13.2234% → fixed QPA
    14.9871% (constraint costs +1.7637 pt) → fixed QPA + η +0.2
    13.9907% (shape gains −0.9964 pt). Net +0.7673 pt → fails.
    QPA freedom (1.76 pt) is worth more than any profile knob can return, so
    this family is structurally unviable while phase fractions stay free.
    The remaining explanation for sample 7 is the structural model (the
    bundled fluorite reference is an ideal CaF₂ CIF while the literature phase is
    an anion-excess defect structure) — an input-data problem, not an engine
    problem.

Quality

  • Test suite: 596 passed / 0 failed / 0 skipped.
  • Cross-checked against MAUD on 7 real samples; mean Rwp 11.030% vs MAUD
    11.203%, every sample within tolerance.

Download

  • XinRefinement-1.7.0-setup.exe — Windows x64 installer (zh/en/ja).

Upgrade notes

  • Registration: the production registration key is embedded in the release
    build, so previously issued registration codes remain valid.
  • Only the installer is published; portable / internal evaluation builds are
    not released.

XinRefinement v1.6.4 — Audit remediation batch 3 (robustness and read performance)

Choose a tag to compare

@XinProf XinProf released this 12 Nov 10:37

XinRefinement v1.6.4 — Audit remediation batch 3 (robustness + read performance)

Baseline v1.6.3 | Theme: the five findings from the 2026-11-11 global audit that
could be fixed without touching any numerical path — bounded recycling of the
project extraction directory, no more silent loss of a corrupt embedded result,
pattern loading reduced from 2–3 reads to 1, repeated e.s.d. evaluation collapsed
into one call, and stale step records no longer accumulating on a reused runner.

Zero numerical change: engine and numerical paths untouched; the golden 50/50
case is still 3.4074156640940005, CLI/API bit-for-bit unchanged.

Download

File Size SHA256
XinRefinement-1.6.4-setup.exe 75.3 MB 258afa309558adfc18ca4a9bed3193213c2982b56006460291fae0a6f5b90701

Release installer (offline registration, zh/en/ja installer UI).
The internal portable build is for internal acceptance only and is not published.

What changed at a glance

v1.6.3 v1.6.4
Temp directory one new directory per project revision under %TEMP%\xinrefinement_projects, never cleaned up bounded recycling: keep the newest 8 / 7 days / 512 MB; the directory currently in use is never removed
Corrupt embedded result result panel silently went blank status bar reports "result/refine.json could not be parsed and was ignored"
Opening a large pattern the same file was read 2–3 times read once (Shimadzu .txt 5 → 1, FullProf .dat 3 → 1)
Save project / export report each recomputed e.s.d. (≈4p full-profile evaluations each, on the main thread) computed once per refinement; failures are cached too
Reusing one runner state.records accumulated, record count doubled cleared before the run, record count equals the step count

Details

A07 · Bounded recycling of the .xinr extraction directory

Extraction directories are still named <stem>_<mtime> and still reused for the same
mtime; recycling runs after each extraction with three limits: newest 8 entries,
7 days, 512 MB total for the set. protect (the directory in use) is the only
absolute guarantee
; keep is a count threshold only, so an entry inside the newest
8 but older than 7 days is still removed. Every OSError during cleanup is swallowed —
cleanup is a side job and must never break opening a project. A recycled project is
simply extracted again the next time it is opened (expected behaviour).

A08 · A corrupt result/refine.json is no longer silent

load_project() now returns a warnings list (the legacy .xrp path exposes the same
key with an empty list), and apply_project() appends each warning to the status bar
instead of showing a modal dialog. The caught exception set is still
UnicodeDecodeError / JSONDecodeError — it was not widened to Exception.

B07 · Single-read loading for legacy text patterns

.xy/.txt/.dat/.csv used to be read 2–3 times (text decode, a 4096-byte header sniff,
then another full read inside the vendored reader). Now the raw bytes are read once and
threaded down through four optional, defaulted arguments — head= / full_text= /
raw= / key= — so default call behaviour is bit-identical. Pattern.meta keys and
values, and the type and text of every parse error, are unchanged (verified by running
the HEAD implementation side by side, not by letting the tests grade themselves).

B02 · e.s.d. computed once instead of three times

MainWindow._esd_for() caches on id(res) + rwp + iters + id(_phases) + phase
names and cells. A failed None is cached as well — the failure path also costs 4p
full-profile evaluations, so retrying it is pointless. new_project() clears the cache.
The arguments and the numbers produced by estimate_esd() are untouched.

records · run_all() clears state.records before running

Re-running a single RecipeRunner accumulated the previous run's StepRecords
(record count doubled, rollback indices shifted). The list is now cleared in place, which
preserves the existing semantics of run_all returning the same list that rollback()
slices.

Upgrade notes

  • Project files: the .xinr format is unchanged; v1.6.3 projects open directly.
    warnings is a new key, and older versions simply ignore unknown keys.
  • Long-term use: stale session directories under %TEMP%\xinrefinement_projects
    older than 7 days or beyond the newest 8 are removed automatically. Your .xinr
    files are never touched.
  • Known limits: when two processes open projects at the same time, the recycling
    in process B may treat process A's session directory as "the 9th oldest" and remove
    it (protect is only effective within its own process); and re-saving a project does
    not carry over the warnings seen when it was opened (a warning is a fact about that
    particular load).

Build and reproducibility

  • Build environment: Python 3.10.11 + PyInstaller 6.22.3. SOURCE_DATE_EPOCH is taken
    from commit 7973d98 (1794479509).
  • The executables are reproducible: same epoch, same source → XinRefinement.exe
    and XinRefinementCLI.exe are byte-identical across builds.
  • The installer is not reproducible: ISCC honours SOURCE_DATE_EPOCH for embedded
    file times, but the PE header keeps the compiler stub timestamp, so two packagings of
    the same input have the same size and different bytes. The published hash is
    therefore the hash of the uploaded artifact
    — do not re-upload to chase the HEAD
    timestamp.
  • The registration key is not in version control. Release builds inject it from
    XR_REG_SECRET or xinrefinement/registration_secret.txt (gitignored) so that
    previously issued registration codes keep validating; without a key the build falls
    back to the community placeholder, which is the path taken by source builds and the
    internal variant.

Verify

sha256sum XinRefinement-1.6.4-setup.exe
# 258afa309558adfc18ca4a9bed3193213c2982b56006460291fae0a6f5b90701

Full regression: 579 passed / 0 failed / 0 skipped (v1.6.3 baseline 553; this
release adds 26 cases).

Note for developers: do not put pytest's --basetemp inside the repository working
tree. tmp_path reuses and prunes numbered subdirectories, which in restricted
environments can trip a bulk-delete guard and raise SystemExit: 1 during fixture
setup, showing up as a wall of bogus ERRORs. Use --basetemp="$TEMP/xinr_...".

XinRefinement v1.6.3 — e.s.d. report value column fix and low-risk audit fixes

Choose a tag to compare

@XinProf XinProf released this 11 Nov 07:46

XinRefinement v1.6.3

Fix: e.s.d. report "Value" column (A11)

  • Root cause: EsdResult did not store the true parameter values, so the "Value" column in text/JSON reports was incorrectly filled with e.s.d.
  • Fix: added a values field (concat(x_final, c), in strict same order as labels); the text table now uses zip(labels, values, esd), and JSON gains values. The e.s.d. / covariance / correlation / content e.s.d. computation paths are all untouched.

Low-risk audit fixes (F01–F06, from the previous round of global code review)

  • F01: the phase display name name_mineral was lost during structure_with_cell / xyz step rebuild → display name degradation
  • F02: legacy _unpack rebuild of ProfileParams omitted size_strain
  • F03: _solve_raw in api.report() omitted refine_po, inconsistent with api.refine()
  • F04: .xinr/.xinm pattern serialization %.10g → %.17g, bit-for-bit identical on save/reopen
  • F05: CLI -o now auto-creates a missing parent directory; to_report texture completion for model/degree/coeffs
  • F06: step buttons still showed Chinese after a language switch → now uses tr(f'step.{key}')

Tests

  • De-flaked the cancellation run test: the cancellation predicate changed from an absolute wall clock >0.15s to a residual checkpoint count, asserting the real contract (cancellation is not recorded as failed, and the unpromised precondition "at least one step completed" is no longer appended)

UI

  • Splash screen shrunk to 800×600, and the version number text removed

Verification

  • Full regression: 547 passed / 0 failed / 6 skipped (the 6 skips are due to environment tier differences — missing COD archive / gemmi / xinmatch on this machine — not a regression)
  • golden CLI Rwp = 3.4074156640940005 bit-for-bit unchanged
  • installer size 75.0 MB, sha256 184695ec1238773e1aedc7547e5b4072bdacadfd52025c968fe77839e3456ae4