v63 — 2026-07-25
A gravidade de São Paulo, não a do livro-texto. O g do modelo passa de 9,81 para 9,7864 m/s² — o valor local, medido por gravimetria absoluta no IAG-USP. Os 9,81 são ≈0,24% altos aqui, e todo MDT que o app carrega e todo passeio contra o qual o modelo foi calibrado ficam na região metropolitana de São Paulo. Resultado: β = m·g/k_eff cai de 0,7585 para 0,7567 kJ/m (veja o leitor de coeficientes em Parâmetros) e toda energia cai ≈0,24% em terreno de subida (≈0,2% em terreno misto). As rotas ótimas quase não mudam — o custo inteiro reescala por um fator praticamente comum; um bundle antigo, recarregado, recalcula com o g novo e lê ≈0,2% abaixo do raster que guardou. A mudança re-sincroniza o app com o modelo do repositório irmão (bicycling-energy-model, Entry 27): enquanto os dois discordavam, o applet de lá lia exatamente essa diferença. É mudança só de constante — nenhum motor mudou, e a paridade bit a bit com o backend Rust segue exata.

v62 — 2026-07-12
Direções ≠ 8 no backend nativo/nuvem. O backend Rust (0.2.0) agora serve o recurso completo de direções de movimento (4–128): conjuntos de Farey, arestas longas integradas ao perfil e carimbo de passagens nas células varridas, em paridade bit a bit com o motor JS (21 casos novos de paridade, todos com delta zero). O app confere a versão do backend (≥ 0.2.0) antes de despachar — um binário antigo ignoraria o campo silenciosamente — e diz o motivo quando o backend é velho demais. A estimativa de tempo modela os termos de memória das direções extras.

v61 — 2026-07-12
Fallback silencioso agora é opcional. Nova opção Falhar em vez de cair pro navegador no painel da nuvem: marcada, uma falha na nuvem (boot da VM, senha, ou o próprio cálculo) aborta o run com o erro na linha de status, em vez de recomputar silenciosamente nos workers do navegador — o fallback é amigável pra runs casuais, mas mascara problemas da nuvem e gasta o laptop num cálculo dimensionado pra 128 vCPUs.

v60 — 2026-07-12
Ergonomia da troca de máquina na nuvem. Timeout do /cloud/start 15→45 s (o redimensionamento roda dentro dessa chamada — ~11 s medidos — e o aborto do cliente virava um falso "orquestrador inacessível"); o religamento no meio do boot leva o machineType selecionado; um aviso de estado indica quando a VM ligada (keep-warm) difere do tamanho selecionado (a seleção só vale no próximo boot); e o novo botão Desligar VM agora força esse próximo boot — trocar de máquina vira desligar → selecionar → calcular.

v59 — 2026-07-12
Nuvem funcionando de verdade — e com escolha de máquina. A CSP agora permite o host real do orquestrador (o modo Nuvem estava silenciosamente bloqueado pela própria CSP desde que foi lançado), a URL padrão aponta pro serviço vivo, e uma URL colada sem esquema ganha https:// automático. Novo seletor Máquina da nuvem (8/32/128 vCPUs, com preço de referência por hora — a VM roda em modo spot, tipicamente 60–90% mais barato): o tamanho é aplicado ao criar/religar a VM e a estimativa de tempo modela o tamanho selecionado (curva de contenção de banda agora sublinear — 128 vCPUs não estimam mais um absurdo 1,4× sobre 8). Cálculos na nuvem dizem "Calculando na nuvem…", e qualquer retorno aos workers do navegador agora diz o motivo na linha de status.

v58 — 2026-07-11
Recortar e reduzir resolução do DEM (1A). Recortar à janela atual recorta o DEM carregado para a janela visível do mapa — aproxime de um bairro num raster enorme (ex. o Sítio Urbano de 135 M células) e fique só com aquela extensão, com cálculos mais rápidos e menos memória (só DEMs geográficos, como os pulls OSM). Reduzir resolução (fator 2/3/4/6×) faz média de blocos N×N ciente da máscara — a mesma transformação do gdalwarp -r average: 6× num raster IGC de 5 m cai em ~30 m, o regime onde o modelo v2 é mais preciso (journal, Entries 19/21); funciona também em DEMs projetados. Ambos passam pelo caminho padrão de carregamento (exportação→reimportação em memória), então herdam tudo: cancelamento de cálculo, máscara nodata, CRS, invalidação de rede/grafo/sonda e a marca de suavização (uma derivação de MDT já suavizado NÃO é re-suavizada), e removem a URL de origem dos metadados de bundle (são derivações locais).

v57 — 2026-07-11
Direções de movimento, string pulling e correção de grade — a nota de pesquisa de conectividade da grade virou opção. Novo seletor "Direções de movimento" (4/8/16/32/64/128; padrão 8 = motor clássico, bit-idêntico): mais direções (escada de Farey, com movimentos longos integrados ao perfil do terreno) reduzem a sobre-estimativa de energia ótima causada pelo serrilhado das rotas (≈⅔ a menos com 16 direções, medido no MDT de SP), a custo de cálculo proporcional. Cálculos de densidade amortizam tabelas de custo pré-computadas por worker; as passagens são carimbadas nas células que os movimentos longos atravessam (corredores contínuos). Valores ≠ 8 calculam no navegador — o backend nativo segue servindo o motor de 8 movimentos. Nova opção "string pulling": encurta pós-cálculo a rota exibida e as top-N (ida-e-volta e maximizar ficam de fora) — a energia mostrada é a soma integrada da própria linha exibida. Novo campo "correção de grade ×c" no bloco de acessibilidade: infla os limiares E₁/E₂ (equivale a deflacionar as energias sobre-estimadas da grade; c* ≈ 1,09–1,12 medido em 8 direções), com aviso explícito de que a garantia de piso conservador deixa de valer; desnecessária com ≥ 16 direções. As três opções persistem nos bundles; a estimativa de tempo acompanha o número de direções.

v56 — 2026-07-11
KPIs de acessibilidade — iniciativa "cidade 300 kJ". Cálculos de densidade agora também amostram o campo de energia de cada referência na célula de todas as outras — uma matriz K×K par-a-par sem custo extra de Dijkstra. Novo bloco "3B. Acessibilidade" nos resultados avalia os dois KPIs ao vivo a partir da matriz em cache: fração da população que alcança ≥ K pessoas com ≤ E₁ kJ (padrão 1 milhão / 200 kJ) e fração que alcança ≥ J% da população com ≤ E₂ kJ (padrão 80% / 600 kJ). A amostragem censitária (IBGE 2022) agora registra a população total M da extensão — cada referência representa M/K pessoas; outras estratégias aceitam um M digitado (o KPI 2 dispensa M). Editar os limiares reavalia na hora, sem recalcular; os marcadores das referências são recoloridos vermelho→amarelo→verde pelo acesso atingido; exportações CSV por referência e da matriz completa. O backend Rust ganha want_matrix no /density (f32×K² anexado, "matrix":K no meta; paridade bit-a-bit com o worker JS, garantida por novos casos +matrix no test-backend.mjs, incluindo regressão de índice com referência descartada — um binário antigo degrada para "KPI indisponível", nunca uma matriz errada). O bloco avisa quando um limiar excede o orçamento de energia do cálculo — ali a matriz é truncada e os KPIs viram um piso.

v55 — 2026-07-07
Suavização do MDT + meta de precisão validada (bicycling-energy-model journal Entry 20). MDTs finos (pixel ≤ 10 m, como os rasters IGC-SP de 5 m) agora recebem uma suavização gaussiana estática (σ = 10 m, normalizada por máscara) aplicada uma vez ao carregar — a mitigação da sobrecarga de resolução do Entry 19, na configuração validada. Novo controle "Suavização do MDT" em 1A (auto / desligada / 10 / 20 / 30 m; o auto pula fontes grossas como FABDEM e re-importações já suavizadas, que o dem.tif exportado agora marca via tag). As alturas são suavizadas no app antes de irem aos motores — paridade JS/Rust intocada; mudar o controle vale no próximo carregamento de MDT. Com esse σ + parâmetros calibrados por ciclista (CdA, Crr, k_s ajustados no próprio histórico — procedimento no journal), o erro validado em passeios não vistos ficou em mediana |Δ%| 3.7 / 2.7 / 4.9 com viés < ±1% em três ciclistas independentes — atingindo a meta de ±5% / ±2%. Nota honesta: a calibração é a alavanca maior; a suavização é o padrão validado que também ameniza o viés sem calibração. Novo test-dem-smoothing.mjs trava a transformação.

v54 — 2026-07-07
Energy-model journal alignment (bicycling-energy-model journal Entries 18/19). Reference-geometry (imported/drawn track) energy readout only — routed-path/field energy is untouched. refEnergyKJ's descent-recovery ε now ports the champion harness's epsGeom estimator exactly: drop-weighted over 30 m-resampled cells, replacing a lumped mean-descent-grade approximation its own comment had claimed but never implemented. Also clarified (comments only, no behaviour change): the v2 engine's trailing descent-cost clamp is provably dead code — a 1.78 M-combo sweep plus 1 402 real rides never triggered it — kept only for JS/Rust bit-parity. New disclosures: the v2 model is tuned for ~30 m DEM sampling, so on the deployed 5 m IGC-SP DTM energies read conservatively high on hilly terrain; and FABDEM is not recommended for energy work on flat/urban terrain, since its per-pixel noise inflates measured ascent.

v53 — 2026-07-03
Icon cache-bust fix. v50's icon redesign kept every filename identical, colliding with the 30-day cache deploy.sh sets on icons/ — a browser or CDN edge that had cached the old bytes wouldn't see the new icon for up to 30 days regardless of redeploys. Every shipped icon file now carries a -v2 suffix, bypassing every existing cache outright. favicon.ico can't be renamed (fixed well-known path), so its cache lifetime is cut from 30 days to 1 instead.

v52 — 2026-07-03
Cross-repo energy-model audit fixes. An audit against bicycling-energy-model (the energy law's spec repo) confirmed v2Edge is correct everywhere it's called, but found graph-mode maximize ignoring mode "to" (always scored forward, unlike top-N and the raster path) and round-mode path energy reporting only the outbound leg instead of the round-trip total. Both fixed, with new regression tests. Removed the legacy QGIS plugins this app was ported from.

v51 — 2026-07-03
Second full-repo review fix batch (see docs/review-2026-07-02-round2-workorder.md). Top-N and length-constrained-DP correctness fixes (true route energy, "até"-mode direction, repulsion clamp); graph-mode spatial-hash rewrite and a T-junction merge fix; bundle geotransform check and a zip-bomb cap; DEM-load and OSM-loader identity guards; worker-pool teardown; export georeferencing fixes; cloud VM lease + boot-order + spoofing hardening; a Content-Security-Policy and help-dialog focus trap; remaining i18n gaps and doc corrections. An adversarial re-review of the batch's own diff caught two more races it had introduced: a water-mask rebuild's staleness check and a DEM load's generation guard could each be fooled by unrelated concurrent activity — both now check the right thing.

v50 — 2026-07-02
New app icon — a yellow ascent arrow crossing a blue topographic contour field, replacing the old bicycle glyph. Regenerated every size (favicon, 192/512, apple-touch, maskable 192/512) from the new design source; same file names throughout, so no manifest changes were needed. Also placed beside the "Simujaules" title in the panel header, which was text-only before.

v49 — 2026-07-02
Full-repo review fix batch (see docs/review-2026-07-01.md). Top-N: an admissible A* heuristic (the old one overcharged gentle descents, so route #1 could cost more than the field's own optimum) and correct route direction in "até" mode. "Maximizar" no longer silently empties under an energy budget. Graph mode: fixed a half-cell terrain/rasterisation/snapping registration error, T-junctions now join the network, chained bridge/tunnel decks flatten as one continuous span, and node snap tolerance is capped at 15 m on coarse DEMs. Native backend: maximize now matches the browser engine under a network constraint, and /single passes are exact on very large DEMs. Hardened a crafted-.gpkg XSS, bundle import now cancels an in-flight compute first, and the compute-time estimate no longer sticks on "estimating…" if calibration fails. Deploy no longer deletes the cloud VM startup script; the service worker installs past the HTTP cache and stops caching large DEM/GeoPackage files. i18n and v1→v2 cost-model doc fixes; the offline census pipeline works again.

v48 — 2026-06-29
Comparação de rotas origem→destino. Com “Comparar com cenário sem rede” ligado, o seletor “Cenário exibido” passa a alternar também a melhor rota, não só o campo: rota no terreno (sem restrição) em azul, rota na rede (restrita) em laranja, as duas juntas na vista de diferença — as mesmas cores da diferença de densidade. Uma melhor rota por cenário. Vale no modo raster (máscara de rede) e no modo grafo (“seguir os vetores”). Passe o mouse ou toque em qualquer uma das rotas para ver a energia e o comprimento das duas de uma vez (mais o Δ — o custo de ficar na rede); funciona com hover no desktop e com toque no celular.

v47 — 2026-06-29
Reverted the MTPI basemaps added in v46 — they were meant for amora, not Simujaules.

v46 — 2026-06-29
Two MTPI (multi-scale topographic position index) basemaps in the basemap dropdown — Pindorama 90 m (COP90, South America, native z10) and Bacia do Paraná 30 m (native z12), XYZ tiles from telhas.pedalhidrografi.co.

v45 — 2026-06-29
Bridge/tunnel deck passes read as a continuous line again. A deck is a 1-cell line over water (unreached by the compute), so its stamped density was faint and dropped by the passes-overlay downsample on big DEMs — a gap along the bridge. The stamp is now dilated by the render stride. Render-only.

v44 — 2026-06-29
v2 energy model. The cost is now physics-based — mass, Crr, CdA, air density, drivetrain efficiency, power on the flat and a climb threshold replace the α/β/η knobs. Per edge it splits rolling from aero (aero charged only off the climbs) with a per-grade descent recovery ε = clamp(min(1, (α/β)/s) − 0.13). Worker, graph engine and the Rust backend move together (bit-parity); reference-geometry energy uses the closed form with a 2 m elevation deadband. Defaults are São Paulo-context (ρ=1.1, Crr=0.008, 75 kg, ~80 W flat); the panel shows the derived α_r / α_a / β coefficients (kJ/m) and a k_smooth knob. See bicycling-energy-model/notas.md.

v43 — 2026-06-26
New Geometria de referência layer: its layer-control row has an ↑ upload button (instead of opacity) that loads a GPX track, drawn as a magenta overlay. Hover or click the track to see its metrics — distance, total energy (the app's α·dist+β·dh model with the current parameters), total ascent and total descent; elevation comes from the GPX or is sampled from the loaded DEM. Also: energy now reads as a plain decimal (e.g. “3357 kJ”) instead of scientific notation. JS/UI only.

v42 — 2026-06-26
Removed the on-map Leaflet attribution strip; the same credits (Leaflet, Leaflet-Geoman, OpenStreetMap, CARTO, Esri, and the pedalhidrografi.co hydrography overlay) now live in the Atribuições / Attributions section above. JS/UI only.

v41 — 2026-06-26
Mobile layout fix (JS/UI only). On phones the locate, hamburger and layer-control buttons now mirror the desktop top-right stack instead of being scattered to the bottom corners, and the layer-control modal opens below them — before, the tall modal covered the bottom-corner buttons so you couldn't tap the layer button to close its own modal. The modal also scrolls vertically when space is tight (its 2-column flow moved into a height:auto child) instead of spilling sideways. Sized for an iPhone SE.

v40 — 2026-06-25
Passes are now always shown as a normalized density rather than raw counts (JS/UI only; display-only and colour-invariant). A single run's passes are subtree-size counts, while the multi-reference density already normalizes by H·W twice — so the two read on very different scales (a 3C channel showing 12 instead of the density's ~1e-10). Count-based passes are scaled by 1/(H·W)² into the same units as the density at every render point; already-normalized fields are left untouched. The colour/pattern is unchanged (a constant scale) — only the numbers and placeholders move to density units. The GeoTIFF export stays counts.

v39 — 2026-06-25
Density-display polish, dataset export/import, and map fixes (JS/UI only; engine untouched). Graph-mode passes: network vectors keep a true orange (terrain a true azure) — intensity is now opacity, not a multiply that muddied mid-values into yellow-brown — and the filtro média N filter applies to the network channel (precise vectors when empty, a rasterised + mean-smoothed field when set; the difference view composites both channels into one orange/azure raster). The 3C.a passes and 3B energy min/max inputs now show the resolved auto numbers instead of the static p10/p90/auto. The sidebar balances its columns evenly once they overflow (was column-1-heavy). The rmsampa-v2 hydrography is crisp on HiDPI displays and no longer vanishes past z17. New in Group 0: export/import of the four inputs in matching formats — DEM and mask as GeoTIFF, network and bridges as GeoPackage. The layer-control panel is resizable to two columns.

v38 — 2026-06-25
A UI/UX review pass (JS/UI only; engine untouched). Units: the budget stays in kJ, so α/β are labelled kJ/m again (reverting v37's J/m); downhill recovery is now a 0–100 % input (default 10), mapped to the engine's 0–1 fraction. The collapse/expand group cue is a quiet grey bar now (was a loud near-white border that out-shouted the status colours); Baixar bundle is secondary so Calcular is the only primary. Drawing shows an armed button state and Esc cancels; the “Concluído em…” pill auto-dismisses; the disabled Calcular button explains what it needs; enabling density auto-opens the points/references group. Accessibility: bilingual accessible names on the icon buttons, file pickers and layer-panel rows, accent-color on native controls, ≥24px touch targets, reduced-motion support. Plus i18n/copy fixes, shorter (non-clipping) select options, and mobile button placement fixes. New cloud example datasets: Viário RMSampa (road network, 1B) and Águas RMSampa (water mask, 1C), covering the full RMSampa region.

v37 — 2026-06-25
UI refinements on top of v36 (JS/UI only; engine untouched). Sidebar groups reordered/renamed (1B–1D, 2A–2C; "maximizar energia" and "origem das referências" removed); the multi-column sidebar fills each column top-to-bottom before the next and scrolls vertically only. The 1B network params and 2A parameters are now 2-col label|input tables (cost coefficients in J/m, budget in kJ). The energy-field / density / legend controls moved into the Controle de camadas panel (its × removed — the layer button toggles it, Esc closes it); Statistics stay in the sidebar. Collapsed groups read grey, expanded white (under the green/orange/yellow/red status colours); 1C/1D go green on drawn geometry alone. The app is now titled Simujaules with a tagline. Fixes: collapsing the sidebar no longer blanks the map; the rmsampa-v2 overlay no longer 404s tiles past zoom 16.

v36 — 2026-06-24
Big UI overhaul (JS/UI only — engine, worker and backend untouched, bit-parity preserved). The sidebar is reorganised into nested, collapsible numbered groups (0 Import/Export · 1 Inputs [1A–1D] · 2 Compute setup [2A–2C] · 3 Results [3A–3D]) with per-group status colours, auto-expand, and a draggable 1–4-column resizer. Layer visibility/opacity/order + basemap (now with an Esri satellite option) + reference markers moved into a non-blocking on-map Controle de camadas panel (floating button, unified per-layer rows). Compute progress/log float in a dismissable bottom pill. Full config persistence + Export/Import/Reset (Group 0), and bundles embed the entire UI state. On-map drawing (Leaflet-Geoman): barrier + passable corridor polygons (1C) and portal lines (1D) feed the compute and round-trip through config/bundles. Smaller: density tweaks follow the displayed channel (terrain vs network), passes auto-max = p90 above auto-min, scientific-notation density bounds, desktop sidebar collapse, locate marker + accuracy circle, credit moved here. Sampling order is now Censo / Sobol / Uniforme (Halton dropped).

v35 — 2026-06-24
App moved to its own domain: now served from the root of simujaules.pedalhidrografi.co (dedicated bucket) instead of telhas.pedalhidrografi.co/simujoules/, and branded Simujaules (an affective typo of joules); absolute URLs repoint to the new root (the RDF vocabulary namespace stays on telhas as a stable identifier, resolved via redirect). And cloud compute (Nuvem — VM orquestrada) now works when the app is accessed remotely: a public, password-gated Cloud Run orchestrator creates/starts/ stops/deletes the GCP VM on demand, and the browser sends the big compute payloads directly to the VM over HTTPS — fill in the Orchestrator URL and the new Cloud password to use it. The VM uses an ephemeral IP and is deleted after 30 days idle (cost → 0). No engine changes.

v32 — 2026-06-24
New IBGE 2022 census (population density) reference-sampling strategy. Instead of spreading references uniformly, it samples them by where people actually live: a live, in-browser port of census/sample_census.py that fetches the census setores inside the DEM from a cloud FlatGeobuf over HTTP Range requests (only the bbox slice transfers, not the ~450 MB national file), weights each setor by population × (area inside the DEM / full area), picks setores by a 1-D Sobol inverse-CDF and drops points inside them with a 2-D Sobol rejection draw. Brazil-only and online-only (like the OSM and cloud features); replaces the current reference set.

v30 — 2026-06-23
Bundles now round-trip the "Comparar com cenário sem rede" (compare) view: the unconstrained energy/passes and the saved difference field are written to the bundle, the compare toggle + e-max-mode are recorded, and load rebuilds the scenario picker and difference view — no recompute. (Graph-mode compare restore is still a follow-up.)

v29 — 2026-06-23
Difference (network-cost) passes view recoloured from red/green to orange (network) + azure-blue (terrain) — additive complements that sum to white (max brightness on overlap) and stay distinguishable under red–green colour-blindness. The two passes control groups now have colour-chip labels (Network / Terrain).

v28 — 2026-06-23
Faster network interpolation on large DEMs: the post-compute IDW fill now sizes its worker pool by deviceMemory instead of a fixed budget that pinned big DEMs to a single worker. Output unchanged.

v27 — 2026-06-23
Cloud keep VM warm between runs toggle: when ticked, a run leaves the VM up to reuse on the next run (only the first run pays the boot); the orchestrator lease + in-VM idle-watchdog stop it after ~15 min idle. Tab-hide no longer stops the VM in keep-warm mode. Default stays stop-after-each-run.

v26 — 2026-06-23
Cloud compute source. The "Use native backend (Rust)" checkbox is now a three-way Compute source selector: Browser (in-page worker pool, the default), Localhost (the native Rust backend, with its URL field), and Cloud. Cloud drives a small local orchestrator (http://127.0.0.1:8079) that boots a pre-baked VM on demand, waits for it to be healthy, runs the compute through it, then stops the VM after the run (and on tab hide). A transfer-size estimate (upload / download / wire time) shows beforehand. Cloud is only available when the applet is served locally (the orchestrator listens on loopback). On any orchestrator or boot failure the run falls back to the in-browser worker pool. Browser/Localhost behaviour and engine bit-parity are unchanged.

v25 — 2026-06-22
Graph mode now RUNS the comparison instead of disabling it. "Constrain to network" is forced on + locked while "Compute on network graph (follow the vectors)" is on (graph mode is inherently network-constrained), but "Compare with unconstrained" stays togglable: graph mode then computes a full-DEM unconstrained raster scenario alongside the graph compute and exposes the energy difference (the cost of being restricted to the network) through the "Displayed scenario" picker (graph / unconstrained / difference). Native backend: "Use native backend (Rust)" moved out of the density panel (always visible) and now also accelerates single-source from/to/round ENERGY fields via a new POST /single endpoint — top-N routes, the destination path and "maximize" stay browser-only (the backend produces no routes), auto-fallback unchanged. energy-worker.js + backend/src/main.rs kept bit-parity (test-backend.mjs +single cases).

v24 — 2026-06-21
Fix: "Calcular sobre o grafo da rede (seguir os vetores)" takes precedence over the raster "Restringir cálculo à rede" / "Comparar com cenário sem rede" toggles, so with graph mode on (it's remembered between sessions) the comparison never ran and its "Cenário exibido" dropdown never appeared. Graph mode now greys out and disables those two toggles (with a tooltip) while it's on, so the precedence is visible.

v23 — 2026-06-21
OSM ele on bridge/tunnel decks (group 1d): "Pull bridges & tunnels from OSM" now also reads node ele tags and uses the mapped deck elevation for the raster portal endpoint heights (per-abutment, with way-ele / nearest-node fallback) instead of guessing from the bare-earth DEM. Unmapped ends fall back to the DEM, so a pull without ele is identical to v22. Persisted in the bundle; engine/backend kept bit-parity. Graph mode stays DEM-based.

v22 — 2026-06-21
New 1c source: "Pull water from OSM" builds the impassable mask straight from Overpass — water areas (natural=water / riverbank / reservoir, incl. multipolygon relations), the open sea/ocean (from natural=coastline, with land-left/water-right orientation), and non-tunnelled waterway=river lines (with a "Rivers (lines) impassable" toggle). Streams and tunnelled waterways stay passable. It reuses the uploaded-mask pipeline (Invert, corridors, overlay, bundle). Needs a geographic DEM.

v21 — 2026-06-21
Fix: "Load FABDEM for current viewport" crashed with "TypeError: t is not a function" — the tile-mosaic loop variable shadowed the i18n t() function after status strings were migrated to it. Renamed the loop variable; the button works again.

v20 — 2026-06-20
Bridge follow-ups. Deck passes/density now show up: a bridge portal jumps abutment→abutment, so deck cells previously rendered no passes even when the bridge carried heavy traffic — they're now painted with the crossing flow. Groups 1c (impassable mask) and 1d (bridges) light up when their data is loaded. And a new 1d "Extract from loaded network" button derives bridge/tunnel decks from the network already loaded in 1b (gpkg tags / OSM other_tags) — offline, no Overpass. Both 1c and 1d now have an "Apply to compute" toggle to enable/disable their effect without clearing the data.

v19 — 2026-06-19
Optional OSM bridges & tunnels (group 1d). A dedicated "Pull bridges & tunnels from OSM" control queries Overpass for bridge (and optionally tunnel) ways over the DEM extent and models each as a level deck between its two ground abutments — useful for inland viaducts over valleys/saddles, not just water (a bare-earth DEM omits the deck, so routing over a bridge otherwise dives into the gap below). Each deck becomes a portal edge between its end cells at the flat-deck cost, relaxed alongside the grid edges, so the route over a bridge and the route under it (a cross-street beneath a viaduct) both stay correct — true multi-level routing on a 2.5-D grid. Graph mode ("follow the vectors") also goes multi-level: the OSM streets pull captures bridge/tunnel/layer tags, so a deck crossing a road at a different layer doesn't connect and deck edges read ~flat. Decks render on the map and persist in bundles as bridges.geojson. The native backend gained matching portal support (bit-for-bit parity). With no bridges loaded, results are unchanged.

v18 — 2026-06-19
Optional impassable mask (group 1c). Upload a binary GeoTIFF (1 = impassable, e.g. water bodies); it's resampled onto the DEM grid by area-coverage majority (≥50% of a cell's footprint impassable ⇒ blocked; outside the mask's extent = passable), so it can have a different extent / resolution / CRS than the DEM. Masked cells block routing. With a vector network loaded (1b), an optional toggle carves narrow passable "bridge" corridors across the mask, each levelled to a smooth profile (land elevation at the shores, ramping to a ±offset at the bridge centre, clamped −5…+15 m). A "show on map" toggle paints blocked water (red) and corridors (green), and bundles gain impassable.tif. Engines unchanged — a run with no mask reproduces prior results exactly.

v17 — 2026-06-19
UI / accessibility pass. The "Load DEM" and "vector network" groups now light up once their data is loaded. Parameter & visualization choices persist across reloads (localStorage; session data like the DEM and points isn't saved — that's what bundles are for). The clear-points / place-random / clear-refs buttons are clustered together, and only the src/dst picker fades in density mode. The sidebar hint paragraphs were removed and their content moved into this Help modal (now also covering maximize, the OSM pull, the native backend, worker sizing and QMC sampling). Accessibility: the status line is an aria-live region, field labels are associated with their inputs, opacity sliders have labels, there's a keyboard focus ring, and the example loaders are buttons.

v16 — 2026-06-19
Time estimate fixed for network/graph runs. Graph mode ("follow the vectors") was estimated with the raster model (~1000× too high — a graph search is ∝ edges, not 135 M cells); it now uses a graph-size model. The IDW interpolation fill is now a separate phase (often the larger one) that scales with the max ray distance, and it's corrected independently of the compute. Toggling interpolate / compute-on-graph / constrain-to-network now updates the estimate. Native backend log line now echoes Emax / mode / type.

v15 — 2026-06-19
Reloading a bundle now restores all saved layers, not just the energy/passes fields. Top-N routes and the maximize path were already in the exported zip (routes.geojson / path.geojson) but the loader never read them back — it redrew the fields and asked you to recompute. They now come back and recolor exactly like a fresh compute, no recompute needed. (Graph-mode edges aren't restored yet.)

v14 — 2026-06-19
Compute-time estimate accuracy fix (was up to ~3× low, now ~±20% with no bias). Backend density on a large DEM was the main culprit: the estimate assumed all cores parallelise, ignoring that the server caps concurrent slices to a memory budget (1-2 fit on a 135 M-cell DEM, not cores-many) — it now replicates that cap (/health reports the memory budget). The calibration probe is now cell-capped (≤~1.5 s on any DEM, anchored on an unsaturated point — fixes a separate up-to-3.8× error on small DEMs), and an online correction learns actual/predicted per engine so the estimate converges to your machine within a run or two.

v13 — 2026-06-18
"Follow the vectors" network-graph mode (new graph-engine.js): with a vector network loaded, an optional toggle routes on the real polyline graph (planarised — junctions at crossings or only at shared endpoints) instead of the rasterised mask, so passes/paths trace the vectors with no staircase; all compute modes; colored-vector overlay; JS-only.

v12 — 2026-06-12
Density worker pool (multi-core) + Dijkstra heap optimisations (~10× density runs). Optional native Rust backend (off by default, density only, auto-fallback). QMC (Sobol/Halton) sampling option for random reference placement. Round-trip budget mode: cap each leg or the round-trip total; round-mode passes count only completable (in-budget) trajectories. "Energy color" passes blend (hue from energy, opacity from passes). Optional vector-network rendering (black lines, configurable ground width + opacity; passes overlay stacks above the network; stacking reorderable via a modal) and a "constrain compute to network" toggle. Basemap selector (OSM / Carto minimal / solid colours). Export of rendered energy/passes PNGs with world files. OSM (Overpass) street-network pull. Constrained-vs-unconstrained comparison: interpolated energy difference + additive red/green passes overlay with per-channel controls (regular and density modes). Interpolation prefilter + worker-pool banding. Density radix-heap engine (in-browser density within ~20% of the native backend on huge DEMs). Compute-time estimate is now budget- and engine-aware (per-DEM calibration probe at load): changing the energy budget, alpha, worker count, or backend now moves the estimate (it previously ignored the budget entirely). Density scales to large workloads: leaner f32 workers + deviceMemory-budgeted pool + opt-in "max compute workers" override; the native backend caps parallel slices to a memory budget (SIMU_MAX_MEM_GB) so it no longer OOM-crashes at high ref counts. Backend liveness/progress fixes. SEO/LLM metadata (description, canonical, Open Graph, JSON-LD, llms.txt, sitemap). Fixes: PWA manifest restored, service-worker install bug, offline caching of CDN libs (SRI + crossorigin), 3-D GeoPackage (Z/M) parsing, stale-result race on DEM/network swap mid-compute, worker-crash handling, XSS hardening of Bundle-before-DEM loading order now restores correctly (pending-bundle re-apply). Density+network Compute no longer aborts silently (null-src re-snap bug); network clicks snap grid-wide (no more snap-radius dead-end); .gpkg geometry column name read from metadata; empty-on-DEM networks rejected loudly.

v11 — 2026-05-13
Wasm engine removed (JS worker only). Fix layered-DP backtrack direction bug (spurious "backtrack_fail"); fix ReferenceError in result metadata.

v8–v10 — 2026-05
Reverse-optimisation (maximize energy) toggle; length-constrained max-cost path via layered DP; DP failures surfaced in the UI instead of the console.

v4–v7 — 2026-05
DEM relief layer (cmocean.phase + slope); locate-me button; OOM fixes for huge DEMs (reservoir-sampled percentiles, stride-capped canvas); drawer/z-index and iOS private-browsing fixes.

v1–v3 — 2026-05-08
First public shell: GeoTIFF DEM loading, asymmetric-cost Dijkstra energy fields, passes count, top-N routes, multi-reference density, vector-network constraint, bundle export (GeoTIFF + JSON-LD), PT/EN, FABDEM viewport loader, PWA shell.