Release notes
What shipped in AdPix, generated from the product changelog.
0.9.1404#
Fixed
- setup-wizardthe contact question offers contact links only
0.9.1399#
Fixed
- insightsa failed follow-up answer can be retried in its own thread
0.9.1397#
Added
- currencya website nobody configured learns what it earns in
Fixed
- wordpressa toman shop reports its real revenue, not zero
- usersa user with no platform link is a 200, not a 404
- techdevices and OS versions the browser really reports (ADR-0130)
0.9.1389#
Fixed
- setupreview round 37 — path-only payment heads, what is paid for, activation returns, quoted title words
- setupreview round 36 — payment returns sure vs maybe, directional cancel promises, PSP list parity
- setup-wizardreview rounds 31-35 — payment returns, cancellations, language mirrors, PSP list
- setup-wizardreview round 30 — cancellations per part, PrestaShop controller keys, product id pairs, checkout order confirmations
- setup-wizardreview round 29 — cancellations across the address, value flags, routed keywords, PrestaShop/Sylius
- setup-wizardreview round 28 — product groups, content presets, OpenCart, flags
- setup-wizardreview round 27 — password pages, cart groups, content presets
- setup-wizardreview round 26 — «پسورد», diacritics, weak «رمز» verbs, products group, content folders never preset
- setup-wizardreview round 25 — «رمز» after a password verb, careers words in group parts, language-home action rule
- setup-wizardreview round 24 — joined password words, careers, «رمز», plain done
- setup-wizardreview round 23 — action keys, failed flags, «رمز», site type
- setup-wizardreview round 22 — plain done pages, «رمز», /products/ by site type, Ghost under a folder
- setup-wizardreview round 21 — home index spellings, order results, Ghost returns
- setup-wizardreview round 20 — language-home routes, AMP twins, narrower done words
- setup-wizardreview round 19 — goal apps, finished orders, language-home routes
- setup-wizardreview round 18 — value-pair cap, keyword reserve, route id-pattern flags
- setup-wizardreview round 17 — canonical page groups, hash-route flags, value pairs
- setup-wizardreview round 16 — hash routes, id patterns, Persian kaf, login/done words
- setup-wizardreview round 15 — hash routes, .php groups, back-test doubt, stale scans
- setup-wizardround 14 review + validator findings
- setup-wizardreview round 13 — language homes, index spellings, Joomla ids, cart/checkout/lead words, option tiers
- setup-wizardreview round 12 — page matching, lead pages on shops, owner ticks
- setup-wizardreview round 11 — query pages, language groups, quiet feeds
- setup-wizardreview round 10 — stopped feeds, page keys, language groups
0.9.1365#
Added
- setup-wizardweb page, admin hub row, report links and docs (ADR-0129 PR4)
- setup-wizardapply, revert and health (ADR-0129 PR 3)
- setup-wizardevidence reads, rule back-test and run routes (ADR-0129 PR 2)
- setup-wizardcontract, ADR-0129, PG 0154 and the pure planner
Fixed
- setup-wizardreview round 9 — one page per spelling, query pages, every-host groups, multi-stream and same-page unticks
- setup-wizardreview round 8 — host-scoped pages, hex-case matchers, honest lead values
- setup-wizardreview round 7 — rescan after live changes, whole-window connectors, web fixes
- setup-wizardreview round 6 — every count of an event covers the same span; a connector is told by its rows
- setup-wizardreview round 5 — the source of an event is decided from certain facts only
- setup-wizardreview round 4 — never act on a guess, say what the bridge line does, two more worlds
- setup-wizardreview round 3 — bridge only forwards what it can, dead and weak sources, needs, safe apply
- setup-wizardreview round 2 on the web page and planner (ADR-0129 PR4)
- setup-wizardreview round 1 on the web page (ADR-0129 PR4)
- setup-wizardrevert never undoes a later hand edit (review round 1)
- ecommerceline items count once per order
- apievent-rule PATCH merges onto the stored rule
0.9.1348#
Fixed
- trendskeep the operator badge text inside the box
0.9.1346#
Added
- didara won deal in Didar becomes revenue on the campaign that earned it
0.9.1342#
Added
- billingtell AdPix Pay which websites are connected
0.9.1339#
Fixed
- webmobile layout fixes for reports, measurement, fraud, consent and onboarding
0.9.1336#
Fixed
- webmobile layout fixes across realtime, AI insights, reports, admin and the date picker
0.9.1331#
Added
- wordpressrecord the login event, not only the identify (1.2.0)
- datalayerlogin and sign_up reach the panel from the dataLayer
Fixed
- weba stale page chunk heals itself instead of a dead-end error screen
- webthe date picker is a bottom sheet on a phone, not a clipped popover
- accessa super-admin's default website is one that exists
- insights,assistant: the AI never links to heatmaps or session replay
- emailinvitations fall back to the inviter's language; add the 0152 down file
0.9.1323#
Fixed
- billingthe upgrade confirmation says what leaves the wallet, and what a short wallet costs later
- realtimea wide screen widens the visitor's identity column, not the dead strip
- invitedan invited trial replaces an unpaid subscription; a paying one is named
0.9.1319#
Added
- emailone branded layout for every email, per-language logo, code-email previews
Fixed
- reportsTotal revenue on Events counts purchases only, never every valued event
0.9.1315#
Added
- trendsevery card has a value + full PDF report by email
- trendsno publishing floor while the network is small
- trendsAdPix Trends — public monthly statistics about Iran's web (ADR-0127)
Fixed
- trendsthe report PDF without blank pages — topics with data only, a real gutter, family labels
- trendsshare previews, Persian badge font, report month, an expert PDF, session epoch, «ادپیکس ترندز»
- trendsPersian pages fully in Vazirmatn + Persian operator names, report mail failures surfaced, no counts
- trendsno change across a counting break (builder 4)
- trendsa rebuild of stale editions runs at once, not at night
- trendshonest operator index, previous period from its own build, no hold on the running month
- trends/ar redirect named Next's internal port (trends.adpix.io:3000)
0.9.1295#
Added
- usersfollow the AdPix Account, and a 30-day bin for people and organizations (ADR-0126)
- billingmirror every website's plan to AdPix Pay (ADR-0125)
- demogive the demo shop a favicon
- demothe sample shop — 25 months of generated data, live visitors and a real demo-shop page (ADR-0124 part 2)
Fixed
- realtimecurrent page is the page the visit is ON, not where it landed
- realtimedecouple card charts, merge the page tables, define the metrics
- demoa demo browser sends no per-person writes (recent website, browser-list import)
- demoa Pay setup error no longer strands demo leads; retry with backoff and a send-again button
0.9.1279#
Added
- demoread-only live demo behind an email link, leads to the Pay CRM (ADR-0124)
- webe-commerce counts and money go compact on every figure
Fixed
- demoleads Pay accepts, a website picker, a working CRM link, demo-aware menus
- webmetric cards stop stacking three-up on a phone
- webthe onboarding stepper reads, and its summary rows line up
- webthe MMM intro says the model starts from 26 weeks, not ~2 years
- docsquote the en mmm-vs-lift FAQ answer so its front matter parses
- buildpip waits 300 s and retries 8 times in the Python job images
- measurementR138 dip wording, long_events span, R133 baseline, fa spacing
- measurementR137 a calendar that fails to build fails the run
Changed
- i18n(fa): plain-space plurals in this branch's Lift strings (brand book)
0.9.1255#
Added
- webpop out the AI assistant into a movable, resizable window
- techone Web Vitals table behind a dimension switcher, operators first
- webAdPix loader + animated dots while the assistant is thinking
Fixed
- measurementR136 a thin OFF state counts its post days in thin_confounded
- measurementR135 a counterfactual that rests on a holiday effect learned from past years reads low confidence
- webe-commerce page fits a phone
- webAI assistant works on a phone
- measurementR133 a frame-start partial occurrence is no past one; model_only lists nothing learned
- measurementR132 the organic estimand takes the learned holiday offset too
- weba shorter scope callout, and a plan badge that names no plan
- measurementR134 0144's test runs the files statement by statement, as migrate.sh does
- measurementR129 failed event columns make the calendar unknown — no day is excused
- measurementR128 a long holiday with no measurable past reads low confidence
- measurementR127 "long" is the calendar row's own length, never the frame's view of it
- measurementR126 R113's holding applies to long rows only; a short row's dip is data as in round 12
- measurementR125 the holiday-confounded bar scales with the event's share of the post window
- measurementR124 a held holiday dip in the pre window the fit does not model reads low confidence
- measurementR123 docs — Nowruz spans 13 days, long holidays read low confidence, WINDOW_DROP_SD, R95–R122
- measurementR122 0144's down reverts only the Nowruz row 0144 itself moved
- measurementR121 the Lift list and detail read low confidence from the result's confidence
- measurementR120 the network monitor and the pooled noise see what a user sees
- measurementR119 a low-confidence Lift result is outside the BH family
- measurementR119 the ON state too: a few pre days of an event never set the level of many post ones
- measurementR118 a state of an event the fit levels on a few pre days is learned from the site's past
- measurementR116 a holiday the pre window cannot estimate is learned from the site's own past occurrences
- webthe API & MCP page says its access is account-level
- apimake the WHMCS and WordPress module downloads public
- measurementR112 web + contract — a low-confidence Lift result states its impact, then why
- measurementR115 realworld.outcome() threads the built action type to v2()
- measurementR114 confidence follows the site's own past occurrences, never an event's name
- measurementR113 a holiday traffic dip is measured with low confidence, not refused traffic_drop
- measurementR112 a holiday-confounded Lift test is measured with low confidence, never refused
- measurementP3 realworld.v2 takes the action type (default 'other', as before)
- measurementP3 the R93 immaterial-event drop uses the fitted pre length
- measurementP3 the validity look-ahead stops at yesterday, never today's partial day
- measurementP2-7 boundary tests for the calendar bridge, traffic_low run, R97 bar, nb_alpha and cut onset
- measurementP2-5 traffic_low is never blind on a short history and never reads the post window
- measurementP2-2 an excused holiday day is still a KPI-break candidate
- measurementP2-1 a calendar row excuses only a suppressed holiday's dip, never a dead tag
- webtranslate the tag installation check message
- measurementR107 a null Lift test around Nowruz is refused or reads null, never an effect
- measurementR111 the short-frame tag-gap test no longer shares the R104u id with the round-12 xfail
- measurementR107 the Nowruz false positive is a named strict xfail, not hidden
- measurementR110 test_adverse_R99 asserts the reason, not only the invalid days
- measurementR109 a carried pause is skipped only when it lasts through the pre window
- measurementR108 the design path's immaterial-event drop skips its own pre window
- measurementR106 a declared stop explains a dip only while organic traffic holds
- measurementR105 a partial recovery at the action is the level coming back
- measurementR104 restore post_returns in tracking_gaps (short frames)
- ingesta withheld identifier never reaches the nightly bridge either
- ingestwithhold a colliding identifier instead of aborting the identify
- ingestone customer's identifiers can live on two rows without failing the identify
- measurementR96 follow-up — robust traffic dispersion and one-day bridges in calendar runs
- measurementa low-traffic refusal says 'traffic_drop', not a tag outage (round-12 P3)
- measurementround-12 P3s — refusal detail in one transaction, unrated currencies of the test window only
- measurementpast_deviation never reads an occurrence inside the test's own pre window (round-12 P3)
- measurementR102 a pause that began before the pre window does not cap a campaign_start
- measurementR101 the tested stop of a traffic-heavy channel is the ads effect, not 'no signal'
- measurementR103 a closure starting on the last post days is judged with the days after it
- measurementR100 R93's drop never touches an event active in the test window
- measurementR99 a closure inside an ads pause is still a KPI break
- measurementR98 the placebo monitor and the design bank drop immaterial events too
- measurementR97 a new level is 'back' only half way to the old one (hysteresis)
- measurementR96 a holiday or a noisy day is not 'no traffic signal'
- measurementR94 a site-wide traffic dip that returns in the post window stays low (found by the grid)
- measurementR91 unrated spend that makes the paid total unknown is named
- measurementR92 a campaign member with unknown channel levels gets no curve prior
- measurementR90 a day with no traffic signal is invalid, not data
- measurementR89 a tag gap's recovery in the post window keeps the run a gap
- measurementR93 an event regressor that never moved the KPI in its own history is dropped
- measurementR88 cost with no FX rate is unknown to activity detection, never a stop
- i18ndrop stale fa overrides that hid the Lift/Incrementality names (PG 0142)
- measurementR87 the ledger's expiry recheck builds without the outage check
- measurementR85 an auto test failed by an error is re-measured once, like an interrupted one
- insights,i18n: no heatmap/replay links in briefs; fa Lift/Incrementality names; Tracking code -> Data stream
- measurementR84 the placebo monitor logs and counts failed builds, and never records an all-failed run
- measurementR83 a raw-named stop is judged against the sources active near the action
- measurementR82 spend with no FX rate is unknown in the paid total, never 0 or left out
- measurementR81 the re-baseline look-ahead stops at the action date
- measurementR80 the stop day's partial-spend row no longer cancels the declared stop
- measurementR79 an outage-check error fails the run, never measures every day valid
- measurementR78 the placebo monitor runs the outage check
Changed
- i18n(fa): drop the English brackets from the Incrementality sidebar label
- measurementdelete R89's post_returns branch in tracking_gaps (R94 subsumes it)
- adminone home per connection item — server keys, tag identifiers, install, sign-up (ADR-0123)
0.9.1146#
Added
- datalayerread GA4 e-commerce from the dataLayer — one checkbox per tracking code, off by default (ADR-0122)
Fixed
- measurementR74 bounded look-ahead, days after it required, never inside the post window
- measurementR76 a declared stop zeroes only what it names, until the next cost row
- measurementR73 an unknown channel-day is unknown; a cut needs every channel known
- gateR72 D1 maps the conflict to this test's bridge row
- gateR72 foundation D1 accepts a recorded calibration_conflict as reaching the fit
- gateR72 verify-measurement-foundation stamps its Lift test v2 like the api
- measurementR71 a sparse KPI's closure with stray orders is never a new level
- measurementR68 only a persistent new level is re-baselined, and only its own days
- measurementR70 a raw channel name reaches the dominant-channel cut rule
- measurementR69 material_cut reads the known spend; an unknown post window is never a cut
- measurementR67 a declared stop is a known 0; unknown channels bound the total
- measurementR66 v1 fit MAPE is reported on the KPI's own scale
- apiR65 a no_effect Explorer channel is not counted as an increment
- measurementR63 the v1 counterfactual point never goes below zero
- webR64 the Lift report never says 'drove' for a non-significant result
- measurementR62 only a material spend cut makes a KPI collapse ambiguous
- measurementR61 traffic_drop_after_spend_cut counts the test's own window
- measurementR60 a day with no cost imported yet is unknown, never a spend cut
- measurementR59 a permanent level shift is re-baselined; a closure never is
- measurementR58 a gap run that began before the spend cut stays a tag outage
- webR57 a not-significant Lift result states its estimate instead of 'Neutral'
- measurementR55 the v1 counterfactual is unbiased on zero-heavy counts
- measurementR56 a channel whose episodes don't exclude zero publishes no measurable effect
- measurementR54 a traffic drop that starts with a spend cut is not a tag outage
- measurementR48 an episode's spend cut keeps only the post days from its date
- gateR53 no measurement gate lets a DB-backed test skip silently
- measurementR48 after a spend cut a KPI collapse is warned about, not masked
- measurementR46 R47 R49 redesign the KPI-break detector around a frozen weekday baseline
- ingestfetch the tracker build deps from the in-country npm mirror
- measurementR52 an auto episode's 2nd interruption invalidates it, counted by member
- measurementR51 an always-late source's launch is judged against its own delay
- measurementR50 a first upload backfilling more than one cadence is not a launch
- weba date picker stops at today, and one time control governs the KPIs
- measurementR44 the L8 gate checks every automatic test over the closure
- ingestinstall the tracker build deps with npm ci
- gateR45 verify-mmm-v2 leaves DB-backed adverse tests to the adverse gate
- measurementR44 a KPI that collapses while the tag keeps firing is invalid, never an effect
- measurementR43 on a masked post window actual_total = counterfactual_total + lift
- measurementR42 a batch source's live start is judged against its own cadence
- measurementR41 the v1 'mixed' halves are scaled like the sd they are compared with
- measurementR40 no pooled bank for a masked post window; pool only full-window tests
- measurementR39 an episode whose auto test is interrupted again and again is invalidated
- measurementR38 a source is stopped on the upload clock alone; its zeros run to the newest row seen
- webacquisition + traffic read as one screen — one toolbar, metric cards, a colour per channel
- measurementR37 only an active arc the MMM uses blocks a newer recommended calibration
- measurementR36 a windowless v1 bridged row is rejected as not an arc, not as an invalid source
- measurementR35 an interrupted auto test expires its episode and is measured again
- measurementR34 one coverage rule — upload clock, live start, every active source of a channel
- measurementR33 activity detection reads every currency at one rate, so a devaluation is not a change
- measurementR32 a run of low days is a tracking gap when its joint Poisson tail is implausible
- measurementR31 Lift scores outage days in the post window as unknown, not as zeros
- measurementR30 Explorer convergence judges ESS on the published decision quantities
- measurementR29 rerun/edit refuse automatic Lift tests (409 auto_test)
- measurementR28 the hourly Explorer tick skips the year of cost reads when nothing changed
- measurementR27 one calibration per change: recommended and the MMM drop overlapping duplicates
- measurementR26 every invalidation deletes the lift-sourced calibration; the MMM checks its source
- measurementR25 one cost-coverage rule per source: first..last row covered, only the tail is unknown
- measurementR24 a candidate whose own pre window already moved is a blend, not a change
- measurementR23 a low day is a tracking gap only when it is implausible at the property's own count level
- measurementR19 a source with fewer than three uploads never reads silence as a stop
- measurementR21 an event's past deviation is the mean of absolute moves
- measurementR20 pooled-noise bucket divides pre_metric by the pre days it summed
- measurementR19 a batch-uploading cost source is not a stop; undetected auto activities are retracted
- measurementR22 re-queue of a test deleted mid-request answers 404, not 500
- measurementR18 a channel past its response curve is held at current, never cut by the grid
- webR15 Lift shows calendar events judged immaterial; docs record rounds R15-R17
- measurementL6 network pooling survives done tests without finished_at
- webR10 the Explorer's fitted/evaluated dates use the website's time zone
- webR8 an arc-only calibration shows no ROI as — and its arc in a Measured change column
- measurementR9 an arc calibration must have ended and have sane magnitudes
- measurementR12 the planner/optimizer trust region always contains current spend
- measurementR7 week by week gives no status to a prior-driven channel
- measurementR2 the Explorer api never emits NaN/Inf (null instead of an empty 200)
- measurementR6 a running Lift test cannot be re-queued (409 test_running)
- measurementR15 a recurring calendar event that never moved the KPI does not cap the grade
- measurementR14 Explorer convergence judged on the reported channels' parameters
- measurementR13 v2 fitted line through the likelihood's mean transform
- measurementR5 an FX refusal skips only its own episode's split
- measurementR4 the split's channel efficiency comes from channel-level members only
- measurementR3 a switched-off single-source channel is a stop, not unknown forever
- measurementR2 no average iCPA at zero spend in the Explorer summary
- measurementR1 Lift fits only on the trailing run of recorded pre days
- measurementI1 the Explorer reads current spend on covered days only
- measurementI7 late-data recheck goes oldest-checked first; expiry withdraws the MMM calibration
- measurementI5 the Explorer is re-evaluated daily at today's spend
- measurementI6 one converged() rule decides whether an Explorer fit is published
- measurementI4 the Explorer models every member channel of a kept episode
- measurementI2 episodes stay 14 days and record each member's exposure
- measurementL2 manual activities are always listed as concurrent
- measurementL7 ledger split reads an MMM prior only from a run on the episode's KPI
- measurementI3 a campaign activity enters the channel curve at channel-level spend
- measurementI1 per-source cost coverage — a lagging importer is unknown, not a stop
- measurementL5 a rerun that no longer calibrates retracts its bridged MMM row
- webD1 MMM calibration note describes the review flow and the ARC form
- webF9 MMM directional verdict says limited evidence, not a ranking guarantee
- webL4 Lift labels lift-times-AOV revenue as an estimate
- webD1 MMM calibration — recommended badge, activate all, ARC manual form, needs-experiment path
- webI5 I6 Incrementality shows when the curves were evaluated; unconverged channels get no action
- webL2 Lift validity shows unestimable calendar events and invalid days
- webL6 Lift calibration labelled own history / pooled network / model only
- measurementD1 MMM calibration UX: recommended flag, activate-recommended, ARC form on POST /mmm/experiments
- measurementL6 one network bank per property x bucket; monitor tests like a real v2 test
- measurementL5 a refusal or failure clears every Lift result and its MMM calibration
- measurementI5 I6 explorer read: evaluated_at, unconverged status passes through, no action for prior-driven
- measurementL5 lift rerun/edit clears every derived result + series/placebo; delete deactivates its MMM calibration
- measurementF5 saved plan's 14-day prediction simulates carryover per draw (mmm.channel_draws.alpha)
- measurementF4 planner + v1 optimizer bounds intersect the trust region (400 invalid_bounds when empty)
- measurementF8 rank-deficient spend columns are not identified
- measurementF6 v2 MMM health and restrictions commit with status done
- measurementF5 persist the per-draw weekly adstock decay in mmm.channel_draws.alpha
- measurementF2 rolling-origin folds see only experiments ended before their cutoff
- measurementF1 fit band and PPC use the fitted likelihood's noise
- measurementL2 list unestimable calendar events and cap the grade at C
- measurementL7 read an MMM prior only from a run whose target is the KPI's
- measurementL4 order value from the test's own pre window, through the checked report
- measurementF3 planner/optimizer current spend converts with the dated core.fx_rates like money.FxTable
- measurementF7 v2 MMM run writes the frozen input snapshot
- measurementL3 date an order by its winning version's local day
- measurementL1 attach the invalid-day mask over the whole Lift frame
0.9.987#
Added
- insightsbriefs that hold up in a meeting — verified numbers, follow-up threads, branded PDF + PPTX (ADR-0120)
Fixed
- webthe visitor snapshot splits this visit from the visitor's history, and reads in the reader's language
- insightsVazirmatn decks, thin waves, no empty charts, legacy-brief notice, glowing ask box
- insightsPersian that reads right, follow-ups on top, branded PDF + deck (ADR-0120 follow-up)
- webdashboard overview reads top-down — range on the KPIs, CTAs in card headers, live as a widget
- speedan unmeasured Web Vital is "—", never a green 0; CLS 0 for a page with no layout shift
- edgepass /pii.js + /vitals.js at the edge and in the Tag-Gateway bundle; CD loads the CHECKED-OUT Caddyfile
0.9.957#
Added
- webADR-0116 Demographics + Tech sections (operator, hidden location, in-app, RFM, devices, versions, page speed)
- audience-techADR-0116 session_tech + web_vitals capture, vitals.js, six audience/tech report endpoints
Fixed
- Codex review findings on phase 4 — comparison after grouping, free orders, stale session rules
- auththe sign-in callback's discovery survives a dropped browser connection
- measurementone order is one purchase, counted once in the website's currency (ADR-0119, F06 + F18)
- identitytwo customers on one browser are two people (ADR-0118, F07)
- opscutover copies a complete snapshot and renames at a quiet moment
- opssort-key cutover in two steps with a performance check; a replicated ClickHouse in the test gate (N1, item 2.2)
- ingesta broken view never fails the raw insert; skipped pushes counted and rebuildable (N3, item 2.3)
- ingestaccepted events are never dropped; bounded buffer with 503 back-pressure (ADR-0116 D1, F01)
- trackerkeep events until the collector answers, and retry without new activity
- ingestserver-side events reserve first and never lose the CRM handoff
- trackerone visitor id when storage throws; testMode travels on the request
- identityone phone rule everywhere — Persian digits, 00 and cc-without-plus identify the same person
- ecommercetwo lines of one order with the same or no item_id stay two lines
- apikey events are matched exactly as configured on every reader
- workerrefuse the Kafka source instead of starting a worker that drops failures
- ingestone bad value can't sink a shared batch; sessions survive deploys; retries don't double conversions
- review findings on phase 4 — stale events, segment scope, order gate
- reportsname what each number counts; sessionless revenue is an Unattributed row (F19)
- ingestdate events when they happened, not when they arrived (F10)
- ecommerceorders, revenue and refunds come from the purchase events (ADR-0119, F16)
- auththe sign-in callback's discovery survives a dropped browser connection
- measurementone order is one purchase, counted once in the website's currency (ADR-0119, F06 + F18)
- identitytwo customers on one browser are two people (ADR-0118, F07)
- opscutover copies a complete snapshot and renames at a quiet moment
- opssort-key cutover in two steps with a performance check; a replicated ClickHouse in the test gate (N1, item 2.2)
- ingesta broken view never fails the raw insert; skipped pushes counted and rebuildable (N3, item 2.3)
- ingestaccepted events are never dropped; bounded buffer with 503 back-pressure (ADR-0116 D1, F01)
- trackerkeep events until the collector answers, and retry without new activity
- ingestserver-side events reserve first and never lose the CRM handoff
- trackerone visitor id when storage throws; testMode travels on the request
- identityone phone rule everywhere — Persian digits, 00 and cc-without-plus identify the same person
- ecommercetwo lines of one order with the same or no item_id stay two lines
- apikey events are matched exactly as configured on every reader
- workerrefuse the Kafka source instead of starting a worker that drops failures
- ingestone bad value can't sink a shared batch; sessions survive deploys; retries don't double conversions
- faPersian digits in every plain {n}, engine-aware ROAS readiness, translated MMM/planner/import reasons, Jalali measurement dates
- faPersian AI insights, 0/0 engagement, tag banner copy, Jalali e-commerce/coverage dates, translated ROAS + cost-import reasons
0.9.0#
Added
- Layered, query-time channel groups — add a channel (e.g. Yektanet) and all matching history reclassifies instantly, no re-ingest.
- Paid vs Organic rollup and a channel-management screen.
- Traffic-source quality report — flags broken UTM tagging (orphan UTMs, malformed mediums, click-id/source mismatches).
- Advanced export (CSV/JSON with metadata, human headers, totals).
- Beta version surface: version badge, What's-new, and feature maturity labels.
- Three-tier RBAC (System/Account/Property) with cascade + data restrictions.
- GA-style Measurement IDs for multi-tenant/multi-site tags; session_start/first_visit + engagement events.
Changed
- Tracker reliability: batching, retry, page-hide flush; ≤4KB.
- Property switcher defaults to your own org.
Fixed
- Unknown Measurement IDs no longer mis-attribute to the default property.