vMPT 1.2.1 — slitlet-aware collision protection + UI cleanup

Changelog

All notable changes to vMPT are recorded here. The format follows
Keep a Changelog and this
project adheres to Semantic Versioning.

[1.2.1] — 2026-06-03

Patch release. Two real bug fixes on top of v1.2.0's collision-
protection feature, plus a substantial UI cleanup that came out of
a hands-on review.

Collision-protection fixes

• Row tolerance is now slitlet-aware. v1.2.0 hard-coded
  |Δs| ≤ 1 between source centres, but a 3-shutter slitlet at
  row s_p already occupies rows {s_p−1, s_p, s_p+1} and the
  user-requested rule is "no other shutter at s_p±2 either." The
  evaluator now computes two tolerances at construction time:
  • Protected slitlet ↔ stuck-open (single shutter):
    |Δs| ≤ half + 1
  • Protected ↔ another slitlet (same slit_length):
    |Δs| ≤ 2·half + 1
    where half = slit_length // 2. So the default slit_length=3
    now correctly forbids stuck-open or other slitlets at rows
    s_p±2. For slit_length=5 the buffer scales up to
    s_p±3 (stuck-open) or s_p±5 (other slitlets).
    SHVAL_S_TOLERANCE = 1 is preserved as the per-individual-
    shutter constant for the live-canvas orange overlap glyph (each
    opened shutter contributes its own ±1 zone, so the visualization
    already paints the correct envelope around a multi-shutter slitlet).
• Advanced settings modal sits above the new config modal.
  When the optimizer config dialog was added in this release the
  Advanced settings card stayed at the same z-index, so opening
  Advanced from inside Configure showed nothing — the config card
  drew on top. Bumped Advanced backdrop / card to z-index 1001 /
  1002.

Pointing-tab UI moved into a dialog

The Pointing tab used to stack 10+ optimizer-config widgets, and the
Run optimization button slid below the fold on any window under
~1200 px tall. The whole block now lives in a centered modal:

• The Pointing tab shows a single primary Open optimizer… button.
• The modal (opt_config_modal_card) contains every optimizer-
  config widget plus Run / Cancel and the live status line.
• The existing progress + results modal flow (opt_modal_card) is
  unchanged after Run is clicked — the config card just dismisses
  itself first.
• Both the optimizer config and the catalog editor modals gained a
  top-right × dismiss button.

Help / status text — context-aware

• The help panel on the right side of the canvas is now collapsed
  by default. The toggle button stays in place; one click on
  Show help restores width to its v1.2.0 size with the Quick
  guide + rotating tip. The figure uses fixed
  frame_width/frame_height so the canvas pixel aspect doesn't
  change when the panel collapses / expands.
• The Method dropdown's three-line Democracy / Meritocracy /
  Hierarchy blurb is hidden by default; an ⓘ What do these mean?
  toggle reveals it on demand. The dropdown's own option labels
  already carry the one-line summary.
• The status line under Run optimization was always reading
  "Load a catalog with priorities, then click Run." even when a
  catalog with priorities was loaded. _refresh_opt_status_div()
  now updates it based on (catalog presence, method, priority /
  weight column availability):
  • no catalog → Load a catalog (Input tab) before running.
  • catalog + Democracy → Ready · N sources.
  • catalog + Meritocracy without weight →
    ⚠ Meritocracy needs a weight column.
  • catalog + Hierarchy without priority →
    ⚠ Hierarchy needs a priority column.

Input / MPT tabs

• All path inputs across the Input + MPT tabs now use a unified
  _wrap_path_picker helper: the path TextInput is hidden
  behind an Edit path toggle when empty, and Browse buttons are
  promoted to primary blue. The path auto-reveals as soon as
  it's populated (by Browse, by autoload, or by typing), so users
  always see what's loaded — only the empty default is hidden.
• The MPT tab is grouped into four sections (Import / Save / Load /
  Export) separated by dashed and solid hr dividers, so the 10+
  widgets feel like coherent blocks instead of one long column.
• Renamed the Setting tab title to Settings (singular →
  plural).

Tests

• tests/test_optimizer_protection.py gains 4 new tests:
  parametrize over N ∈ {1, 3, 5} to pin the cached tolerances,
  and a regression that an N=3 slitlet drops at least as many
  unprotected sources as N=1 under an H grating. 139 passed, 4
  skipped in total (up from 135 / 4 in v1.2.0).

[1.2.0] — 2026-06-03

Feature release: shutter collision protection in the optimizer.

Same-row sources on the same NIRSpec detector half (Q1/Q3 → NRS1,
Q2/Q4 → NRS2) disperse onto overlapping detector pixels when their
V2 separation is smaller than the spectrum's V2 half-extent
(app.wavelengths.v2_overlap_distance — 35″ for PRISM, ~500″ for
the H gratings). Until now the optimizer counted both members of
every such pair as observable; the live canvas already painted the
loser orange, but the score didn't reflect that downstream
penalty. v1.2.0 wires the same collision check into the optimizer's
per-pointing scoring so the user can mark high-priority targets as
protected and have the optimizer steer them into rows free of
collisions.

Optimizer core (app/optimizer.py)

• PointingEvaluator accepts new keyword args: protect_mask,
  priorities, weights, disperser, filt, reason. When
  protect_mask is None (the default), behaviour is identical to
  v1.1.1 — the existing 16 optimizer tests are unchanged.
• A new method evaluate_with_stats(...) returns the existing
  3-tuple plus the count of sources dropped by the collision rules
  at this pointing. evaluate(...) still returns the 3-tuple but
  its detected mask is now the kept mask (post-drop) when
  protection is configured, so callers that score via det.sum()
  pick up collision filtering for free.
• Three rules, applied in order at every pointing:
  1. Protected ↔ stuck-open — a protected source landing on a
     row colliding with any shutter flagged as stuck-open (REASON
     == 2 in the CRDS msaoper file) is dropped. Stuck-opens
     always disperse light onto the detector regardless of which
     slitlets the user opens, so the protected target's spectrum is
     unavoidably contaminated.
  2. Protected ↔ protected — within each colliding cluster the
     lowest-priority-number source wins. Ties on priority break on
     higher weight; ties on weight break on lower index (stable).
     Losers are dropped; winners continue to provide collision
     pressure on the next rule.
  3. Protected ↔ unprotected — every unprotected source whose
     row collides with any still-kept protected source is dropped.
• Dropped protected sources do not propagate collision pressure
  to rules 2/3 — if a high-priority spectrum is already
  contaminated we won't compound the loss by also blocking
  unprotected sources in the same row.

Pointing-tab UI (app/main.py)

• New "Protect spectra from collision" group in the optimizer
  sidebar (just below the existing Priority cutoff input):
  • Checkbox: Enable collision protection.
  • Radio: By priority ≤ | By weight ≥ (mutually exclusive).
  • Threshold text input.
  • Live status line: e.g. "12 protected · 240 other (G140H /
    F100LP · V2 overlap ≈ 500″)" — updates as you toggle the
    checkbox, switch the radio, type a threshold, or change the
    current Disperser/Filter.
• _rebuild_merged_catalog now propagates weight when multiple
  catalogs are stacked — previously single-catalog mode kept weight
  (it pointed at the original Catalog object) but merged mode
  dropped it, so the multi-catalog "By weight ≥" rule would have
  silently selected zero sources.

Results modal

• When protection is enabled the Score cell gains a −K
  suffix where K = number of collision-dropped sources at this
  pointing. The Score column is widened by ~36 px so the suffix
  doesn't ellipsis-truncate.
• The hover top-10 prefixes protected sources with 🛡 so the
  user can verify which sources are providing collision pressure.
  A trailing line in the tooltip explains the −K count.
• Header summary line picks up a "🛡 collision protection ON"
  badge with a one-line explainer.

Tests

• New tests/test_optimizer_protection.py — 11 tests covering
  backwards compatibility, input validation, all three drop rules,
  cross-detector-half non-collision, stuck-open handling, and the
  invariant that protection can only reduce a pointing's score
  (never increase it). 3 tests skip gracefully when synthetic
  sources don't happen to land in the geometry the test exercises.
• Existing test suite unchanged: 135 passed, 4 skipped total
  (up from 124/1).

Notes / known limitations

• For the H gratings the V2 overlap distance is ~500″ — comparable
  to the full MSA — so even one protected target rules out a large
  fraction of co-observable sources. The result is physically
  truthful, not a bug; the modal shows the lower kept count so
  expectations match reality.
• Unprotected sources whose rows collide with a stuck-open shutter
  are NOT dropped from scoring (only protected sources have the
  contamination penalty applied to them). This matches the
  user-requested semantic: protection is a high-priority-only
  feature, not a universal contamination filter.

1.1.1 — 2026-06-03

Patch release. Polish + several real bugs in the v1.1.0 optimizer
and catalog editor. The big change is that Hierarchy mode actually
optimises lower tiers now, plus a much richer results table.

Optimizer

• Slitlet centre is now right under the target. The optimizer's
  axy_to_shutter returns 0-based fractional indices, but
  _add_slitlet expects 1-based — the missing +1 was opening every
  slitlet one row up and one column to the left of the target. Now
  centred correctly.
• Confirm dialog before Apply. Clicking Apply #N opens a
  browser confirm: "This will CLEAR all previously open shutters and
  replace them with the optimizer's slitlets." OK → clears + applies
  (single Undo step); Cancel → no-op. Wired via Button.js_on_click
  → hidden trigger TextInput → Python handler. The trigger pattern
  is needed because CustomJS.args only accepts Bokeh Model
  instances, not floats — that's why the Apply button silently did
  nothing in v1.1.0; embed per-button scalars via Python f-string
  interpolation into the JS body.
• Hierarchy mode now genuinely optimises every priority tier.
  Previously DE refinement used weights = 1 at top tier, 0 elsewhere, so DE happily slid to any pointing that kept the
  top-tier count even if it lost lower-tier sources in the process.
  DE now uses auto-derived lex weights (smallest int weights
  such that any higher tier strictly outweighs the sum of all lower
  tiers); their sum is a lex-equivalent scalar that DE maximises
  without violating priority ordering. The grid + multi-stage filter
  phase is unchanged.
• Results table shows tier breakdown. For Hierarchy, the Score
  column reads e.g. P0:4 · P1:12 · P2:30 (46) — per-tier source
  count + total in parens. For Meritocracy, Σw 287.0 (46). For
  Democracy, just the count 46.
• Hover any Score cell to see the top 10 placed sources at that
  pointing, sorted by priority ascending then weight descending —
  IDs + P + W per line.
• Modal widened to 740 px to fit the new columns; Score column
  width is method-specific; cells now overflow: hidden + white-space: nowrap + text-overflow: ellipsis so a label that
  overruns its column truncates instead of wrapping under the row.

Catalog editor

• Numeric sort on Priority + Weight. Both columns are now stored
  as floats with NaN for missing (was strings → "10" < "2"
  lexicographically). An HTMLTemplateFormatter renders the cell as
  a rounded integer or blank; cell edits via StringEditor are
  coerced back to float in _on_cat_edit_data_change so the column
  stays a sortable numeric.
• After a header click, the table scrolls to row 1. Document-
  level click delegate on .slick-header-column resets the table's
  .slick-viewport.scrollTop to 0 (with an 80 ms delay so the
  re-render finishes first).
• CSV save uses _fmt_int_or_blank for Priority + Weight so
  the output is 5 not 5.0 and blanks stay blank.
• Compute w from p / Compute p from w write floats to the
  source (was strings) so the new column stays numerically sortable.

Misc

• The compute_weights_from_priorities helper now correctly
  satisfies BOTH w(p) > w(p+1) AND N(p)·w(p) > N(p+1)·w(p+1)
  using max(w_prev + 1, n_prev * w_prev // n_q + 1) as the
  smallest integer that dominates the prior class (regression-tested
  in tests/test_catalog_ops.py).
• Loader: empty cells in numeric columns now properly become NaN
  even when the source column was masked-int (previously came
  through as 0).
• A couple of additional patterns added to .gitignore so stray
  personal files in the repo root can't accidentally be staged.

Tests

• 124 passing, 1 skipped (same as v1.1.0; no test regressions).

1.1.0 — 2026-06-02

Headline feature: a complete MSA pointing optimizer with three
methods (Democracy / Meritocracy / Hierarchy), plus an editable,
sortable catalog editor. Several quality-of-life improvements
elsewhere.

MSA pointing optimizer

• New panel at the bottom of the Pointing tab. Searches over
  (ΔRA, ΔDec, ΔPA) for an (RA, Dec, V3 PA) that maximises a
  user-selectable objective. Re-implemented in vMPT style from
  hMPT (Eisenstein, McCarty,
  Wu; CfA / Harvard); see app/optimizer.py for attribution +
  algorithm notes.
• Three methods:
  • Democracy — raw source count; ignores priority and weight.
  • Meritocracy — sum of weight of placed sources (MPT-style).
    Requires a populated weight column.
  • Hierarchy — strict priority-tier lex ordering (eMPT-style).
    Multi-stage filter: a higher-priority source is never traded for
    any number of lower-priority sources.
• Pop-up modal with an animated striped progress bar, a spinning
  ring, and a status line showing the current phase
  (Grid: 5,200 / 20,000 · 4.2s elapsed · ~12s left,
  Hierarchy filter: tier 2 / 4 (p=1) — survivors: 18,
  Refining top 10: 3 / 10 · 7.4s elapsed).
• Results table with the top-10 distinct solutions (near-
  duplicates collapsed). Each row pairs score + (ΔRA, ΔDec, ΔPA)
  with an Apply #N button.
• Apply #N sets the pointing AND opens an N-shutter slitlet
  (N from the Setting tab) at every observable target's shutter,
  auto-tagged with the catalog source ID. One Undo step reverts
  the whole apply.
• ΔX = 0 freezes the axis — set ΔPA = 0 to search RA/Dec only
  at the current roll, etc. Both the grid sweep and the DE
  refinement honour the freeze.
• Advanced settings… modal exposes grid resolution (n_RA, n_Dec,
  n_PA), DE max iterations, objective (count/flux), source σ, and
  the APT DVA θ.

Catalog editor

• New Edit catalog… button in the Input tab opens a sortable,
  in-cell-editable spreadsheet pop-up.
  • Single-click any cell to edit. Tab / Enter commits; Esc cancels.
  • Drag inside a cell to highlight text; Cmd/Ctrl-C / Cmd/Ctrl-V
    copy / paste — a custom capture-phase keydown handler bypasses
    SlickGrid's column-copy default so only the selected text is
    copied.
  • 🗑️ icon at the end of each row deletes that row.
  • ↶ Undo / ↷ Redo for every edit, delete, derivation, and
    column add (100-step history).
• Column picker — toggle which columns are visible. Extras
  columns from the source CSV/FITS (the loader now preserves every
  column it didn't claim) live alongside the standard set and can
  be turned on or off.
• Add a custom column via a text input + button. Empty by
  default; useful for weight, reference, etc. Round-trips
  through Apply changes and Save as CSV.
• Compute w from p and Compute p from w buttons derive
  one column from the other:
  • w(lowest p) = 1; for each higher-priority class, the smallest
    integer w(p) satisfying w(p) > w(p+1) AND
    N(p) * w(p) > N(p+1) * w(p+1). Guarantees strict-dominance:
    one source at any tier outweighs every source at all lower
    tiers combined.
  • p from w groups unique weights descending and assigns
    priorities 1, 2, 3, …
• Save as CSV with a Browse… file picker.
• Apply changes & close commits the working copy to the
  in-memory catalog so the eMPT bundle export reflects edits.

Catalog model

• Catalog.weight is now a first-class field (sibling of
  priority). Loader detects weight / w / wt / weights
  aliases. Empty cells in numeric columns properly become NaN
  (previously masked integer columns silently became 0).
• Extras columns the loader didn't claim are preserved on the
  Catalog.extras dict (object arrays, original column name as
  key) and surfaced through the editor's column picker.

UI polish

• run.sh gained --port N, --fits PATH, --jpg PATH,
  --wcs PATH, --catalog PATH (repeatable) flags. Mutual-
  exclusion rules: --jpg and --wcs come as a pair; --fits is
  exclusive with them.
• Tabs renamed: Image → Input, Aim → Pointing, Pick → Setting.
• Pointing tab now also hosts Disperser/Filter (was on the
  former Pick tab). RA/Dec inputs share a row; V3 PA/APA share a
  row; Visibility date + button share a row.
• Canvas pixel aspect locked — frame_width / frame_height
  match the loaded image's pixel W:H exactly. Window resizes
  letterbox around the canvas; the image is never stretched.
• Sequenced autoload: run.sh --jpg ... --wcs ... --catalog ...
  loads the image first and the catalogs strictly after, via an
  on_complete callback chain so the catalog overlay never races
  the image's _set_image_and_recenter.
• Status bar moved out of the scrollable sidebar column and
  pinned to the bottom-left of the viewport (position:fixed) so
  it can't render on top of tab content.
• Optimizer Advanced settings moved into a pop-up modal.
• 6 new tips in the help-panel carousel — run.sh args,
  optimizer, catalog editor, multi-catalog, pixel aspect, big-ID
  mod.

Tests

• 124 passing (was 96 at v1.0.1). New coverage: catalog weight
  column, mod-1e7 + empty-cell NaN handling, optimizer correctness
  (radec→Axy, quadrant inverse, centration monotonicity,
  Hierarchy-vs-Democracy divergence, dΔ=0 freezes, dedup), the two
  weight↔priority compute helpers in app/catalog_ops.py.

1.0.1 — 2026-05-21

Patch release. Two large quality-of-life corrections — accurate
per-shutter wavelength values, and a much friendlier catalog
loader — plus polish on the catalog UI and overlay defaults.

Wavelength accuracy

• Per-shutter dispersion table for every (disperser, filter)
  combo, derived from numerical integration of the pipeline
  reference files via spacetelescope/msaviz.
  Lives at data/dispersion_cutoffs.npz (19 MB compressed) and is
  regenerated by scripts/precompute_dispersion_cutoffs.py. Replaces
  the old linear V2-shift approximation that was wrong in two ways
  for PRISM:
  • Previous PRISM gap was held at 2.7–3.2 μm everywhere. Real
    gap location varies dramatically across the MSA (5–95 %
    spread: gap_lo 0.65–3.59 μm, gap_hi 3.03–5.02 μm) because
    PRISM dispersion is highly non-linear. The new lookup gives
    msaviz-accurate values per shutter.
  • PRISM endpoints used to drift with V2; in reality they're
    essentially constant (msaviz spread is ~0.01 μm).
• Q3 / Q4 PRISM shutters correctly report "no gap on this
  spectrum" — their spectra fall entirely on one detector.
• Grating endpoints updated to match the pipeline-reference
  sci_range instead of the slightly narrower JDox "useful range":
  • G140M/F100LP: 0.97–1.89 (was 0.97–1.84)
  • G140H/F100LP: 0.97–1.89 (was 0.97–1.84)
  • G235M/F170LP: 1.66–3.17 (was 1.66–3.07)
  • G235H/F170LP: 1.66–3.17 (was 1.66–3.07)
  • G395M/F290LP: 2.87–5.27 (was 2.87–5.14)
  • G395H/F290LP: 2.87–5.27 (was 2.87–5.14)
  • G140H/F070LP: 0.70–1.27 (was 0.81–1.27)
• vMPT does NOT depend on msaviz at runtime — only the
  precompute script does. The shipped npz is everything the app
  needs.

Catalog loader: looser column matching

• _norm() lowercases, strips bracketed/parenthesised unit
  annotations ([deg], (deg)), collapses non-alphanumerics, and
  peels trailing unit/epoch tokens (deg, degrees, rad,
  arcsec, J2000, ICRS, FK5). All of these now match RA:
  RA, ra, RA[deg], RA(deg), RA_deg, RAJ2000,
  Right Ascension, ALPHA_J2000, R.A.[deg]. Same for Dec
  (including Vizier's DEJ2000).
• ID resolution accepts the usual aliases (id, no,
  source_id, objid, srcid, …) plus permissive fallbacks
  (name, label, tag, target, #) — fallbacks honoured only
  when values coerce to integer.
• Missing ID column → synthesised sequential IDs 1..N so the
  catalog still loads.
• Numeric IDs ≥ 10⁷ are taken mod 10⁷ (ID_MOD = 10_000_000)
  so JADES-style 8–9-digit IDs collapse to APT's compact space.
• Priority class strings (P0, P1, …) and masked numeric
  cells now flow through cleanly — the old loader threw
  ValueError on a P0 priority cell and produced 0.0 for
  masked mag / z instead of NaN.

Multi-catalog

• Load multiple catalogs at once. Each gets a colour chip in
  the sidebar list. Toggle visibility per-catalog with a checkbox;
  × to remove; ▲ / ▼ to reorder the visual stack.
• Per-catalog marker colours cycle through an 8-entry palette
  (yellow / magenta / pale green / coral / lavender / sky-blue /
  white / salmon), picked to read clearly on dark fields and avoid
  the other overlay colours.
• Z-order by list order (earlier-loaded catalogs draw on top)
  with alpha decay by depth (1.0 → 0.35 floor). Matched-shutter
  targets always render fully opaque so a "picked" marker is never
  visually demoted.
• Sessions serialise the list (catalog_paths) — workspace JSON
  remembers each path and its enabled flag.

MPT-importable catalog (export)

• Output is now a superset of the input catalog — every input
  source is included, plus any synthesised entries for slitlets
  without a real match. The Label column carries real or
  vMPT_synth so downstream tools can tell which is which.
• Integer IDs only, extracted as the largest digit run from
  the original token (so RJ0600-10274-P0 → 10274). The original
  string token is preserved in the Label column for traceability.

Overlay defaults

• Operable-shutter stroke: 0.75 px → 1.0 px.
• Spectral-overlap fill alpha: 0.10 → 0.20.
• Spectral-overlap edge colour now explicitly orange (#d97a00)
  — when you reveal the edge via the stroke slider it now matches
  the orange fill instead of Bokeh's default blue-grey.

Tests

• 96 passing (was 63 at v1.0.0). Coverage growth concentrated on
  the catalog loader (column aliases, ID synth, mod-10⁷, string
  IDs, name-as-numeric-ID) and the wavelength model (every
  disperser × filter combo verified against the msaviz table).

1.0.0 — 2026-05-20

First public release. The tool is feature-complete for hand-picking
JWST/NIRSpec MSA shutter configurations on a target field and
exporting a bundle that loads into APT MPT and the eMPT pipeline.

Highlights

• Interactive shutter picker with N-shutter slitlets (N ∈ {1, 2, 3, 5}),
  snap-to-nearest-operable, undo / clear, double-click highlights,
  shift-click to move the pointing, wheel-zoom and pan.
• Live overlays — MSA outline, operable shutters (silver edge),
  stuck-open (dark-red outline), user picks (red fill), spectral
  conflicts (orange fill, stackable), 5 fixed slits (gold), catalog
  targets (yellow / green when matched), lime pointing cross.
• APT-ready bundle export — 6 files per export, with role-prefixed
  filenames (MPT_*, vMPT_*, eMPT_*). The MPT plan JSON matches
  APT's reference schema field-for-field; the <catalog>.cat uses
  JDox-recognized column names (ID, RA, DEC, Weight, Primary, Label).
  Labels distinguish real catalog rows from vMPT_synth synthesised
  entries.
• APT plan importer — load any MPT_plan.json, shutter mask CSV,
  local .aptx archive, or fetch by JWST program ID directly from
  STScI. Reads multi-plan archives (e.g. program 1208 with 40+ plans).
• Bundle round-trip — Save session → load session restores
  pointing, V3 PA, disperser/filter, every open shutter with its
  target_id + role, the highlighted set, and the image + sidecar
  paths. Point at either MPT_plan.json OR vMPT_workspace.json —
  the sibling auto-loads.
• Responsive layout — canvas stretches to fill the browser
  window; sidebar / help panel scroll on overflow; left-sidebar
  fixed at 340 px, right help panel at 340 px.
• Rotating tip card in the help panel (13 hand-written tips,
  15-second rotation with CSS fade-in).
• GitHub version-check on startup — non-blocking background
  thread compares the local HEAD to origin/main; shows a
  dismissible amber notification if the local copy is behind.
• Custom favicon (4 MSA quadrants + lime pointing cross).
• One-page summary slide generator (build_vmpt_slide.js,
  pptxgenjs-based).

Science correctness

• MSA geometry sourced from pysiaf (NRS_FULL_MSA); 138.575°
  intra-MSA rotation, V2/V3 reference at (378.563, −428.403).
• APA = V3 PA + V3IdlYAngle (mod 360); both quantities are
  surfaced in the status bar and editable from the Aim tab.
• Operability read from CRDS jwst_nirspec_msaoper_*.json —
  failed-open shutters always disperse and contribute to the
  spec-overlap calculation.
• Spectral overlap — |Δs| ≤ 1 cross-quadrant via NRS1 (Q1↔Q3)
  and NRS2 (Q2↔Q4) detector pairing; per-grating V2 half-extent
  (PRISM 35″, M-gratings 200″, H-gratings 500″).
• Wavelength endpoints per disperser+filter, clamped to the
  grating's intrinsic range (no spurious PRISM > 5.3 µm tooltips).
• Source matching uses APT's Unconstrained Source Centering
  rule (full shutter pitch including bars).
• WCS Jacobian uses astropy.SkyCoord.spherical_offsets_to —
  cos(Dec) factor handled correctly at non-equatorial fields.

Example data shipped

• example_a370/ (43 MB) — JWST NIRCam F182M+F200W+F210M FITS of
  Abell 370, target catalog, GTO-1208 APT MPT plan, shutter-mask CSV.
• example_r0600/ (21 MB) — JWST NIRCam F090W+F200W+F444W JPG of
  RXCJ0600 + WCS sidecar + 28k-source target catalog. JPG re-encoded
  at quality 85 (was 251 MB) without changing WCS.

Tests

• 63 tests, ~5 s. Run with pytest tests/.
• Coverage: session bundle round-trip, MPT plan parser (incl. .aptx
  archives), eMPT format byte-compatibility, MPT catalog writer
  format guard, wavelength model, image loaders, end-to-end export.

Known limitations

• plannerSpecification block in MPT_plan.json carries sensible
  defaults (matching APT's reference schema) but its dither /
  search-grid parameters don't reflect any vMPT internal state —
  APT uses them only as starting values for re-planning.
• Bokeh single-session state: opening the same server in two
  browser tabs lets picks bleed across them. Use one tab per user.
• Older pysiaf PRD (PRDOPSSOC-068) lags the online version by
  ~0.05″ for some apertures; safe to ignore unless you need
  milli-arcsec geometry.

Acknowledgements

Export-bundle format calibrated against eMPT
(Bonaventura et al. 2023, A&A 672, A40). Coordinate plumbing builds
on pysiaf (NIRSpec apertures) and astropy.wcs. Visibility
windows queried via jwst_gtvt.
MPT catalog and plan JSON schemas follow the
JDox MPT documentation.