Sweep of remaining swallowed-error sites flagged by an audit pass:
- weather/gefs_client.ex: byte-range download task exits log url + reason
before being surfaced as {:error, _} (matches the HRRR fix).
- live/path_live.ex: terrain-profile load failure, native-duct lookup
failure, and ProfilesFile.read failure now log warnings with path
endpoints / midpoint / valid_time. Previously rendered a degraded
result silently.
- live/rover_live.ex: station-mutation {:error, _} now logs id + reason
instead of dropping into {:noreply, socket}.
- application.ex: the boot-time backfill_missing_home_qth Task.start is
now wrapped in try/rescue. A DB issue at boot would have crashed the
task silently (no Logger handler attached pre-Logger crash format).
GefsFetch silently dropped crashed score-computation tasks via
`{:exit, _reason} -> []`, hiding partial-grid outages from k8s logs.
HRRR range downloads mapped exits to `{:error, reason}` without logging
the URL, so the failing range was lost by the time the caller surfaced
the error.
All three sites now Logger.error with enough context (fh, valid_time,
url) to debug from production logs.
Adds Microwaveprop.Radio.TextSanitizer to remove BOM, zero-width
spaces/joiners, C0 control bytes (except CR/LF/TAB), and DEL, and to
collapse NBSP and other unicode whitespace into a regular space.
CSV: applied to the whole upload at the top of preview/2.
ADIF: applied per-field-value after extraction so the byte-counted
length tags still slice the right substring.
Adds a checkbox in the rover sidebar that constrains the suggested
candidate spots to lat/lon points tagged as :ideal in /rover-locations.
When enabled, Compute.run replaces grid cells with synthetic cells at
each ideal location's exact coordinates, inheriting the propagation
score from the nearest grid cell so atmospheric forecast still factors
in. Path-clearance, terrain, building, and canopy penalties run
unchanged on those snapped candidates.
New globally-shared list of rover-friendly (or off-limits) parking
spots. Anyone can view; logged-in users can add new locations and
edit/delete their own. Each location stores lat/lon, free-text notes,
and a status enum (:ideal | :off_limits).
* Microwaveprop.Rover.Location schema + migration
* Rover.list_locations/0, create_location/2, update_location/3,
delete_location/2 (mutations require User; ownership-scoped)
* /rover-locations LiveView in the public live_session — Add button
is disabled for anonymous visitors, Edit/Delete buttons render
only on the user's own entries
* Tests for context + LiveView covering anonymous read, owner-only
edit/delete, and form submission
Rover ranking now applies a per-cell tree-canopy clutter penalty
(1 dB / 8 m, capped at 4 dB) so cells in dense forests get down-ranked
even when their per-station path clearance survives. PathTerrain
already accounts for trees ON the link path; this surfaces "I'm in a
forest, every direction is foliage."
Canopy.BulkFetch downloads Lang et al. 2020 10 m global canopy-height
3-degree COGs and slices each into 1° × 1° uint8 .canopy tiles via
gdal_translate, the runtime image now ships gdal-bin. Run on prod via:
bin/microwaveprop rpc 'Microwaveprop.Canopy.BulkFetch.run(32.8, -97.0, 200)'
Tile cache lives under /data/canopy/. Already-sliced tiles are skipped.
Adds Microwaveprop.Canopy module that reads 1° × 1° uint8 per-degree
canopy-height tiles (mirrors SRTM .hgt naming/layout, 30 m resolution).
Returns 0 m for areas with no tile on disk so the lookup is safe to
thread through the existing path-clearance pipeline before any data
lands.
Wires canopy into:
- PathTerrain.obstacle_top: per-sample blocker = ground +
max(building_m, canopy_m), so the rover ranks paths through forests
as obstructed even when no buildings sit on the line
- CandidateDetail profile: each sample now carries canopy_m alongside
building_m
- Rover-detail SVG: green "trees" polygon stacked on terrain, under
the red building polygon
- Path calculator elevation chart: green Tree Canopy dataset between
terrain and buildings
Data prep (download Potapov 2019 / Lang 2022 GeoTIFF, slice to per-degree
uint8 tiles, drop in /data/canopy/) is a separate one-shot operation —
without tiles, lookups return 0 and the existing behavior is unchanged.
Compute.run now accepts a `progress` callback and reports human-readable
labels before each pipeline step (loading grid, looking up elevation,
checking road access, scoring cells, etc.). RoverLive feeds it via
send/2 to itself and renders the latest label as small text to the left
of the Calculate button while it's running.
Each profile sample now carries the tallest building height within 80m
of the path point. The candidate-detail SVG renders these as a red
polygon stacked on top of the terrain so blockage from buildings is
visible alongside ridges. max_obstacle_m now reflects building tops too,
so the clearance label downgrades when buildings sit on the link.
Adds a per-cell building-clutter penalty so the algorithm down-ranks
spots in built environments where buildings on multiple sides
scatter/block signal regardless of which station you're aiming at.
Penalty = max_height_within_75m / 5 dB, capped at 6 dB. Path-clearance
already accounts for buildings ON the link path; this is the
"surrounded by stuff" signal.
Phase 2/3/5 of the building-blockage support:
* Parser streams csv.gz tiles into compact records (centroid, radius,
height) — keeps RAM ~32 B per polygon, drops -1.0-height entries.
* Index buckets records into 0.01° (~1 km) grid cells in a public ETS
table for concurrent reads; max_height_near/3 and records_near/3
scan only the 9 nearest buckets.
* Loader lazily parses any cached quadkey before a Calculate so the
scorer is terrain-only when tiles aren't on disk yet.
* Rover.PathTerrain now treats each path-sample obstacle as
terrain_elev + max_local_building_height, so blocked LOS through
recent construction actually reduces clearance.
* Selecting a candidate pushes a candidate_buildings event with the
buildings near each rover→station path; the hook renders them as
height-coloured circles (yellow <10 m, orange 10-30 m, red >=30 m)
with a tooltip showing height in meters.
Mix tasks don't run in production, so the orchestration lives in
Microwaveprop.Buildings.BulkFetch — callable from an attached iex on
the prod pod via:
iex> Microwaveprop.Buildings.BulkFetch.run(32.8, -97.0, 500)
Default 500 mi radius around DFW yields ~440 candidate quadkeys at
zoom 9; only those present in the MS dataset index (typically a few
hundred over CONUS) get downloaded. Concurrency 4, per-task timeout
3 min, results logged with downloaded/cached/failed/missing counts.
The mix.task is kept as a dev-only thin wrapper.
Phase 1 of building-blockage support: implements quadkey math (Bing/MS
tile encoding), looks up the MS dataset-links index for UnitedStates,
and downloads csv.gz quadkey tiles to /data/buildings (overridable via
:buildings_cache_dir). Index is cached 24h; tiles are written once and
reused. No path-analysis wiring yet — that's the next slice.
Prod logs showed road_proximity took 186 s on a single Calculate.
Req's receive_timeout only catches TCP idle, not slow streaming, so a
trickling Overpass response could not be cancelled. Wrap the call in a
Task.yield/shutdown deadline and cache successful results by 0.05°-rounded
bbox so adjacent Calculates skip Overpass entirely.
- Detail panel now reads as an aiming card: "Aim 220° (SW) · clearance +12 ft"
with the precise bearing in degrees, plus rover/obstacle elev on a second line
- Elevation profile labels swapped from min/max-of-profile to start (rover) and
end (station) elevations, since their positions in the chart now match
- Map draws a dashed indigo polyline from the candidate to each selected
station when a candidate is opened; cleared when the panel closes
Calculate occasionally hung in production for ~30s. Adds per-phase
timing logs to Compute.run so the slow step is visible in pod logs,
and trims the Overpass road-proximity request to a 10s server-side
timeout / 12s receive timeout (no Req retries) so a slow road API
fails soft to no road penalty instead of stalling Calculate.
- Local prominence: each cell's elevation vs. its 8-neighbor ring
(~1.3 km radius) feeds a small additive bonus so broad hilltops
rank above isolated SRTM voxels
- Road proximity: one Overpass call per Calculate fetches drivable
ways inside the bbox; cells beyond ~0.5 km of any road get a
light dB penalty, gated off in tests
For each (rover cell, fixed station) pair, sample SRTM along the
great-circle path and add a per-link dB bonus proportional to the
rover's elevation above the highest intermediate terrain. Rover spots
on hilltops with clear sight to multiple stations now rise to the top.
Also vendor Leaflet's layers control PNGs and copy them under
priv/static/assets/css/images/ on build so /rover stops 404'ing on
layers-2x.png.
The /path calculator showed "0 / 9 HRRR points" in production despite
the on-disk profile store being current. Root cause: profile_from_cell/2
treated the cell's :profile key as a wrapper sub-map and called
Map.put_new on it — but the :profile key actually holds the vertical
pressure-level LIST. Every point sample crashed with BadMapError, the
crash propagated as {:exit, _} through Task.async_stream, and the
consumer silently dropped all 9 results.
Fix: stop wrapping. Cells are already flat HrrrProfile-shaped maps;
just stamp lat/lon (from the caller, since cells don't carry their
own coords — those are the map key) and valid_time onto the cell.
Audit + log every other async error path so the next silent failure
isn't invisible:
- PathLive HRRR point lookup
- Propagation.point_forecast per-hour reads
- Viewshed ray crashes
- IemClient ASOS network fetches
- RtmaClient range-download tasks
- Recalibrator factor-vector batches (positive + negative samples)
- MapLive forecast preload tasks
- RoverLive station resolution
LiveViews already had handle_async/3 exit clauses with logging. The
gap was always in Task.async_stream consumers that wrote {:exit, _} -> []
without surfacing the reason.
Add the rule to CLAUDE.md and project memory so this never repeats.
Also fix a pre-existing skewt_svg.ex compiler warning where
@critical_label_min_dy was used before being defined.
The 18-hour forecast chart on /path now shows the same weighted-
criteria panel the /map popup builds when the user clicks a grid
cell. Hovering a dot fires a server roundtrip that fetches
Propagation.point_detail at the path's midpoint for that valid_time;
results are cached client-side so re-hovering is instant.
`Propagation.point_detail/4` now returns a `:profile_source` tag —
`:exact`, `{:fallback, fallback_valid_time}`, or `:unavailable` —
distinguishing real per-hour breakdowns (Rust pipeline writes a
profile file for every f00..f18 step) from rare cycle-miss fallbacks
where we approximate from a recent analysis profile. Stale comment
about "only f00 persists profiles" updated.
Map utilities (FACTOR_META, FACTOR_ORDER, factorBar,
factorExplanation, scoreTier, FactorMeta, ScoreTier) are now exported
from propagation_map_hook so the new path hook reuses the exact same
look + copy.
The grid-cached profile is trimmed to 13 levels (1000→700 mb) for
memory reasons, so the skew-T plot showed nothing above 700 mb.
HrrrClient.fetch_profile/4 now accepts a forecast_hour opt; SkewtLive
issues a per-point fetch against the same cycle the cached row came
from to get the full 1000→100 mb set, falling back to the cached
truncated cell on any failure.
Time selection no longer round-trips through the URL — `select_time`
updates state in-process and triggers a focused :load_skewt_time
async, so refreshing or sharing a /skewt?q=... link always lands on
the most recent analysis hour.
Three additions that bring the diagram closer to the SPC sounding
viewer's full-information chart:
* Wet-bulb temperature trace, plotted as a thinner blue line under
the dewpoint trace. Comes from a Stipanuk-style bracket between
the level's dewpoint and dry-bulb temperatures (already used by
`SkewtParams` for the wet-bulb-zero level — now public so the
renderer can call it per level).
* Environment virtual-temperature trace and parcel virtual-T
trajectory, both as thin pinkish dashed lines. The parcel
virtual-T is the actual buoyancy curve CAPE integrates against,
so showing it next to the dry-bulb parcel ascent makes the
moisture correction visible. New `SkewtParams.virtual_temp_c/2`
and `/3` helpers, plus `SkewtParams.mixing_ratio/2`.
* Critical-level rail on the right edge. Each of LCL / LFC / EL /
0 °C / WBZ that resolves on the current profile gets a short
horizontal tick at the right of the plot and a `<label> <p> mb`
tag just outside it — matches the SPC chart's `143 mb (mix)` /
`158 mb (mix)` annotation column. SkewtLive computes the
pressures (and converts the height-keyed 0 °C and WBZ from
SkewtParams back to pressure via `pressure_at_height/2`) before
handing them to `SkewtSvg.render/2`.
Wind-derived overlays (hodograph, wind barbs) remain off — HRRR
profile data is wind-free above 10 m AGL.
mix test: 27 / 27 across the four /skewt-related test files,
2,902 / 2,902 + 221 properties on the full suite. mix credo --strict
clean (one cyclomatic-complexity refactor extracted along the way).
When the on-disk ProfilesFile is empty SkewtLive falls back to the
hrrr_profiles Postgres table, but per-QSO HrrrFetchWorker rows
sometimes land with a partial profile (4–12 levels) when individual
byte-range fetches fail. The fallback was happily picking those rows,
which rendered a stubby Skew-T missing the upper troposphere.
HrrrProfileLookup.{list_valid_times_near, read_point_near}/3 now
filter on `coalesce(cardinality(profile), 0) >= min_profile_levels`,
defaulting to 13 (HRRR's standard pressure-level set: 1000/925/850/
700/500/400/300/250/200/150/100 mb plus the two boundary surfaces).
Pass `min_profile_levels: 0` to keep the partial rows for diagnostic
queries.
Confirmed against the dev mirror: the recent table has ~3.1 M rows
at 13 levels and ~250 k at 4–12 levels; the new default skips the
partial slice cleanly.
Tests: 8/8 HrrrProfileLookupTest cases (added two new ones for the
threshold). mix credo --strict clean.
Two changes that bring /skewt visually and analytically into parity
with the NOAA SPC sounding viewer
(https://www.spc.noaa.gov/exper/soundings/):
NOAA-style line set
-------------------
Updated the SkewtSvg style block to match the public SPC chart:
* Temperature trace ........ pure red (#ff0000) at 2.5 px stroke
* Dewpoint trace ........... pure green (#00aa00) at 2.5 px stroke
* Isotherms ................ dashed teal (#14b8a6)
* Dry adiabats ............. solid orange (#f59e0b)
* Moist (saturated) adiabats dashed sky-blue (#38bdf8)
* Mixing-ratio lines ....... dotted darker green (#16a34a), with
labels at 700 mb showing the g/kg
value, drawn only below 600 mb (matches
the SPC chart's lower-band number row)
* Parcel ascent ............ new dashed grey (#475569) line, drawn
whenever SkewtParams.derive returns a
non-empty parcel_trace
The mixing-ratio set is the SPC standard (0.4 / 0.7 / 1 / 2 / 3 / 5 /
8 / 12 / 16 / 20 g/kg) instead of the prior arbitrary set.
SkewtParams: SPC parcel-theory + indices module
-----------------------------------------------
New `Microwaveprop.Weather.SkewtParams` derives the parameter set that
ships in the SPC `*.txt` companion file:
* LCL pressure / temperature / height (Bolton 1980 eq 22)
* LFC pressure / height
* EL pressure / height
* SBCAPE, SBCIN (J/kg, virtual-temperature integration with
surface parcel)
* 3 km CAPE
* Lifted Index (parcel ↔ environment T at 500 mb)
* Freezing level (T = 0 °C height) and Wet-Bulb Zero
* DCAPE (downdraft CAPE — picks the min-θₑ source level in
400–700 mb, descends moist-adiabatically to surface)
* Layer lapse rates: sfc–3 km, 700–500 mb, 850–500 mb
* `parcel_trace` — pressure/temperature trajectory used by
SkewtSvg to draw the dashed parcel-ascent overlay
Wind-derived indices (SRH, BWD, Bunkers motion, SCP, STP, SHIP) are
deliberately *not* produced and the LiveView surfaces a one-line note
explaining why: HRRR persists wind only at 10 m AGL, so per-pressure-
level shear can't be computed from this data source.
LiveView panel restructured into grouped sections (Surface, Convective
parcel, Levels, Lapse rates, Moisture/stability, Refractivity) so the
layout matches the SPC viewer's order. The Levels rows show pressure
*and* height ("872 mb · 1,140 m") for quick cross-reference between
the diagram axis and the readout.
Tests: 8/8 new SkewtParamsTest cases (LCL Bolton check, saturated-air
edge case, full field-set presence, summer profile produces SBCAPE > 0
and LI < 0, capped profile produces SBCAPE = 0 and LI > 0, freezing
level interpolated correctly, sfc–3 km lapse rate, empty profile no
crash). 25/25 across all skewt-related test files. mix test green
(2,902 / 2,902 + 221 properties), mix credo --strict clean, mix
assets.build clean.
Two follow-ups to the earlier /skewt work:
1. Interactive crosshair (matches the SPC-viewer feel from the
reference screenshot). Mousing over the diagram now drops a
horizontal pressure cursor with a top-left readout box showing the
pressure at the cursor's altitude (mb), the height in m and ft,
and the linearly interpolated temperature and dewpoint at that
level. Two coloured dots track the T and Td traces along the
cursor so the user can read the spread visually.
Mechanics: SkewtSvg.geometry/0 exposes the plot geometry (left,
right, top, bottom, p_bottom, p_top, t_min, t_max, skew) as JSON;
SkewtSvg emits an initially-hidden <g class="skewt-cursor"> layer.
A new Skewt hook in assets/js/app.ts inverts pressure_y to recover
pressure from the cursor's viewBox-Y, walks the (descending-pres-
sorted) profile to interp T/Td/height, and updates the elements.
The hook also handles mouseleave to hide the cursor and uses the
SVG's screen-CTM so the math stays correct under any container
aspect-ratio resize.
2. DB fallback for `available_valid_times`/`load_profile`. When the
on-disk `ProfilesFile` has nothing for the location-and-window
(true on dev, transient on prod between chain runs), SkewtLive
now falls back to `HrrrProfileLookup.{list_valid_times_near,
read_point_near}/3` against the `hrrr_profiles` Postgres table.
The lookup uses the same ±0.07° spatial tolerance and the
`(lat, lon, valid_time)` unique index that `Weather.find_nearest_
hrrr/3` already relies on, so the queries are index-only.
Returned shape matches `ProfilesFile.read_point/3` so SkewtLive
reads from either source transparently. NaiveDateTime → UTC
normalisation is centralised in `ensure_utc/1` so the LiveView
sees `%DateTime{time_zone: "Etc/UTC"}` regardless of which side
produced the value.
Tests: 6/6 new HrrrProfileLookupTest cases (list/limit/tolerance,
read nearest cell, nil for far cells, picks closest among matches),
3/3 SkewtLiveTest still green, 2,894 total Elixir tests + 221
properties green via mix test (one pre-existing flaky test in
admin/contact_edit_live re-runs clean). mix credo --strict clean.
mix assets.build clean.
Both retired in light of the 2026-04-25 revision report
(`docs/algo-reports/2026-04-25-algo-revisions.md`):
* HPBL boundary-layer multiplier in `Scorer.score_refractivity/{3,4,5}`
is gone. On the n=47,418 10 GHz HRRR-matched corpus rho_hpbl is
+0.004; binned distance is flat to within ±5 % across 200–2,000 m.
The previously reported 2.3× ratio between shallow and deep HPBL
bins was a small-corpus artefact that disappeared once the matched
corpus passed ~5,000 contacts.
* Native-profile 1.15× duct boost is gone. At 10 GHz on n=52,341
matched contacts, cells where `best_duct_band_ghz` ≥ band ran
*shorter* than no-duct cells (198 km vs 211 km, n=173 vs 44,658).
The Bulk Richardson gate that existed only to suppress that boost
in turbulent conditions is also retired; arity preserved on every
`score_refractivity` overload so callers compile unchanged.
Same retirements applied in the Rust port (`rust/prop_grid_rs/src/scorer.rs`)
to keep the Elixir/Rust parity tests green.
algo.md updated end-to-end:
* Finding 4 (HPBL) rewritten to "no usable signal" with the
n=47,418 bin table.
* Finding 5 (gradient) tightened: load-bearing only at 24 GHz,
set band-specific gradient weight to 0 outside [10, 47] GHz on
the next derive_band_weights run.
* Finding 8 (time-of-day) augmented with the robust hour-bucketed
amplitude index (40 / 82 / 101 % at 10 / 24 / 47 GHz) and a
selection-bias caveat for VHF/UHF (where 129–148 % amplitude
reflects evening contest scheduling, not propagation).
* Finding 9 (sounding) refreshed against the n=27,058 corpus.
* Part 2c "Native-profile duct boost" rewritten as
"retired 2026-04-25" with the falsifying join.
* Part 2c "Commercial-link inverse sensor" promoted from deferred
with the 22-day, 38k-sample SNMP corpus, the +0.61 / −0.61 PWAT
and pressure correlations against the 37-hour HRRR overlap, and
the 04–05 / 10–12 local two-peak morning fade window.
* Part 2b "Pressure" annotated with the U-shape at 10 GHz and the
next-iteration target.
* Part 2d "What stayed, what left" + "Open items" updated to
reflect the four code/doc moves.
119 Elixir scorer tests green, 134 Rust scorer tests green, 2,888
total Elixir tests + 221 properties green via `mix precommit`.
Each fix is covered by a regression test that fails on `main` and
passes on this commit.
Round 1 (initial review):
* propagation: thread `latitude` into the conditions map so
`score_season/4` actually picks up regional multipliers
* hrrr_client / fetcher.rs: `nearest_hrrr_hour` rounds DOWN, never at
a future cycle that NOAA hasn't published yet
* radio: spherical-vector great-circle midpoint replaces the
arithmetic mean — anti-meridian paths no longer fold to Greenwich
* weather: `reconcile_weather_statuses` scales the longitude band by
`1 / cos(lat)` so the bbox stays ~150 km wide at every latitude
* radio/maidenhead: clamp 90°/180° below the field-bucket overflow so
`from_latlon` never emits invalid characters like 'S'
* prop_grid_rs/pipeline: merge HRRR + NEXRAD-derived rain rates and
read `best_duct_freq_ghz` into `best_duct_band_ghz` so the Native
Duct Boost actually fires
* propagation/region (Elixir + Rust): inclusive upper bounds so points
exactly at lat_max get the regional multiplier
* weather/sounding_params (Elixir + Rust): drop the 10 m gradient
floor so HRRR's thin near-surface layers stop hiding sharp ducts
* weather/sounding_params: when the profile ends inside a duct,
finalize it with the highest sample as the top instead of throwing
it away (Rust port already correct)
Round 2 (post-fix sweep):
* radio + commercial: single canonical haversine in Radio (atan2
form); Commercial delegates instead of carrying a second copy that
could disagree at threshold distances
* prop_grid_rs/profiles_file: `snap_coords` matches Elixir's
step-aware snap (`round(coord/0.125) * 0.125`, then 3-dp round) so
Rust-keyed and Elixir-keyed profile maps land on the same cell
* weather/grib2/wgrib2: `parse_lon_val_segment` uses `Float.parse`
uniformly — wgrib2 dropping the trailing `.0` from a longitude no
longer crashes the whole chain step
`apply_edit_to_contact/2` is the path for both owner direct-edits and
admin reviewed-edits. The public contact-map + total-count + gzipped
controller payload caches are built from `private == false` rows, but
only the insert path was clearing them — every other update left
/api/contacts/map and /contacts/map serving the previous snapshot for
up to 10 minutes. A `private: false -> true` flip on a contact already
present in the cached payload was a temporary privacy leak; grid
corrections, callsign changes, and flagged_invalid flips silently
diverged from the live row.
Extracts the three `Cache.invalidate` calls into
`invalidate_contact_map_caches/0` and runs it from both
`apply_edit_to_contact/2` and the existing insert path so the public
view stays in sync.
The /contacts/:id edit form always submits the `private` checkbox, but
`current_value/2` had no `"private"` clause and fell through to the
catch-all `nil`. The Contact schema defaults `private: false`, so an
unchanged box compared `false != nil` and was always tagged as a
change — for non-owners that meant a spurious pending review queued
on every save, and for owners/admins it triggered a no-op direct
update.
Adds the missing clause so the diff sees boolean-vs-boolean and
correctly returns `:no_changes` when the box is untouched.
The /contacts/:id edit form pre-fills qso_timestamp via
`Calendar.strftime(ts, "%Y-%m-%d %H:%M")` and submits whatever the
user typed back unchanged. Two failure modes resulted:
1. `diff_against_contact/2` compared that string against the stored
`%DateTime{}`, never matched, and treated every untouched submit
as a modification.
2. `build_contact_changes/2` then ran `DateTime.from_iso8601/1` on
the space-separated, no-seconds, no-Z form and crashed with
`MatchError` on `{:error, :invalid_format}` — direct owner/admin
submits raised instead of saving.
Adds `normalize_timestamp_field/2` to `normalize_proposed/1` that
parses both the form's `"YYYY-MM-DD HH:MM[:SS]"` and the strict
`"YYYY-MM-DDTHH:MM:SSZ"` shapes into a UTC `%DateTime{}` (with
`NaiveDateTime.from_iso8601` doing the heavy lifting after a small
seconds-padding pass). Empty strings drop the field; unparseable
input drops it too so a typo becomes "no changes" rather than a
500.
