Recount Changelog

[v1.19.2] (2026-09-29) - Tabs and a split view on the main window

New

• Tabs on the main window. A row of folder tabs sits on top of the main window. Each tab is one display mode, labelled with a short name ("DD", "DPS", "Thr"; initials or the first four characters for a mode without one, cut on whole UTF-8 characters). Clicking a tab is SetMainWindowMode, so the window itself is unchanged. The title menu, the arrows, Alt+scroll and the keybindings all change the tab you are on, and its label follows. + adds a tab (starting on the first mode), with no limit. When the tabs do not fit, < and > appear at the ends and scroll the row, as does the mouse wheel over it; selecting a mode brings its tab back into view. Right-click a tab for Close (never the last one) and Move Left / Move Right. Tabs are saved per profile in MainWindow.Tabs by mode LABEL, not index, because indices move when a tracker module is switched on or off; LoadMainWindowData now follows the active tab's label for the same reason. An existing player starts with one tab showing the mode they already had. The window's clamp rect counts the tab row, so dragging the window to the top of the screen cannot push the tabs off it. Show Tabs in the window settings (MainWindow.ShowTabs, default on) hides the row. Designed with the operator on 2026-09-29 from a player's request to watch two displays at once; the split view below builds on them. The tabs are the game's own chat-tab art (Interface\ChatFrame\ChatFrameTab-BG* / -Selected*, used by FloatingChatFrame.xml in every flavour's tree), tinted with the window's title colour so they read as folder tabs of the red title bar and follow the colour settings; the selected tab is lit and solid, the others dimmer. They are drawn one frame level BELOW the window (re-applied whenever WindowOrder.lua raises it) and seated 9 units below the frame's top edge, so the red title bar covers their bottom edge and they read as folder tabs of it; the window's own hit rect is inset by the same 9 while tabs show, so a click on a tab's lower part reaches the tab rather than dragging the window. The 9 was tuned by eye from in-game screenshots (the first build floated above the bar, the second sat on top of it) and is not measured at a known UI scale. Drawing a child one level below its parent is Blizzard's own pattern (FloatingChatFrame.lua:1580), and a positive top clamp inset is how the chat frame takes in its own tab (FloatingChatFrame.lua:70), both per Peer Review. Seen on Classic Era by the operator on 2026-09-29.

• Split view: a second display under the first. Right-click a tab and choose Show Below, and that tab's display fills a second set of bars in the lower part of the window, under a thin title-coloured divider that names it. Drag the divider to move the split (not on a locked window); the drag stops by itself once the left button is up (IsMouseButtonDown, as Blizzard's chat-tab drag does), so releasing with the cursor off the divider cannot leave the split following the mouse, per Peer Review, whether or not the client delivers OnMouseUp there (not verified in a client). Right-click it, or the tab, for Hide Below. Off until a tab is picked, so the window looks as it did. The window's bar drawing now works per pane: a pane is Recount.MainWindow itself (so every existing reader of Rows, DispTableSorted, GetData and the rest is unchanged) or Recount.MainWindow.Lower, a table with the same fields, its own scrollbar (Recount_MainWindow_LowerScrollBar), bars (Recount_MainWindow_LowerBar<n>) and total bar. RefreshMainWindow's body is now RenderPane, run once per pane; ResizeMainWindow lays both out, keeping at least one data bar (plus the total bar when shown) per pane, and drops the split while the window is too short for two panes without forgetting it. The file-local MaxValue is now per pane. The lower pane re-reads its tab's mode by label on every refresh, like the tabs, and clears its tables when that mode or the fight changes. Clicking a bar opens the detail window, graph, realtime graph or bar menu on THAT pane's display, and the detail window's mode arrows keep to it (MainWindow.DetailPane). Report, the title menu, the arrows, Alt+scroll and the keybindings still act on the selected tab (the top pane). Retail and Forever draw both panes from the damage meter (Tracker_Midnight.lua's renderer now takes a pane, through Recount:MainWindowLowerPane). Saved per profile in MainWindow.Tabs.below (a tab index, kept right when tabs close or move; closing that tab turns the split off) and .splitRatio (default 0.5). The Show Scrollbar setting covers both scrollbars. ResetData now clears every pane through FullRefreshMainWindow; it had cleared only the top pane's tables, so the lower pane kept showing the reset players (their records still carry fight data), found by the session's self-audit and pinned by a spec that fails with the old two lines. UpdateBarTextColors now walks every created bar of both panes rather than 0..RowsCreated, which skipped every bar once the total bar was switched on after the others existed (creating bar 0 set RowsCreated to 0). Seen working by the operator on 2026-09-29 on Classic Era (Damage Done over Damage Taken) and on WoW Forever; retail Midnight not yet seen.

• Display mode: "Interrupted" -- who got interrupted, and by whom. The interrupt tracker recorded only the interrupter (Interrupts, InterruptData); the receiving side of SPELL_INTERRUPT ran only housekeeping, so "who keeps interrupting me?" could not be answered. AddInterruptData now also adds Interrupted and InterruptedBy (interrupter -> interrupted cast) to the victim, as Dispels keeps Dispelled / WhoDispelled, and InitFightData sets both up. Starts recording from this version. Like Interrupts and Dispels, it is not carried by LazySync (which sends damage, healing and active time only), so a grouped player sees only what their own combat log recorded. From the recovered self-raised audit finding of 2026-08-18 (inbox 1ec972ac). Its "CC Broken On" half is not built: the CC-break event's receiving side is the crowd-controlled mob, not the player who cast the CC, and the event never names that caster. So the mode would list mobs and still could not answer "who broke my Polymorph".

Changed

• LibDBIcon-1.0 is now a required dependency instead of a bundled copy, on every flavour except WoW Forever. The six non-Forever TOCs add it to ## Dependencies, drop the libs\LibDataBroker-1.1 and libs\LibDBIcon-1.0 loads, and drop both from X-Embeds; .pkgmeta lists libdbicon-1-0 under required-dependencies, so the CurseForge app installs it. The standalone brings LibDataBroker-1.1 with it (its own embeds.xml). A bundled copy only updated when Recount re-released; the standalone updates for every player. Recount_Camelot.toc keeps loading the bundled copies, because the standalone's TOC does not list Forever's Interface 16001, so the files stay under libs/. No Lua changed. Known cost: a player who installs by hand must install LibDBIcon-1.0 too. Not yet run in a client. Requested by TOGTools on the user's 2026-09-25 decision.
• Every Recount tooltip is placed by one function, Recount:AnchorTooltip (Widgets.lua), so none covers the thing it describes: below the hovered frame in the top half of the screen, above it in the bottom half, and aligned to its left or right edge in the outer thirds. It is the rule the minimap launcher already used (a copy of LibDBIcon-1.0's getAnchors), moved out of GUI_Minimap.lua and applied to the main window's bars, title, Compare button and tabs, the settings window's icon tooltips, and the compare graph's controls. The compare graph's value readout still follows the cursor, because it reports the point under the pointer. Asked for by the operator after a tab tooltip covered the tab row.
• The addon-list title is the fleet's orange (|cffFF8000Recount|r) in all seven TOCs.

Fixed

• UnitIsTrivial is called only if the client has it (Recount.lua, where a hostile mob is classed Trivial / Boss / Nontrivial). It is documented in the live and Forever trees only; the Classic trees' UnitDocumentation.lua does not list it. Whether the Classic clients have it anyway is not verified, and Recount has called it unguarded there for years, so this may change nothing in practice; without it a mob is Nontrivial. Every flavour.
• Recount:GetModeIndex no longer reads L[modestring], which made AceLocale report a missing entry for any label that is not a locale key (every mode label on a non-English client). It reads the locale with rawget. Found while auditing the tabs. Nothing in the addon calls it (the keybindings use SetMainWindowModeByLabel, whose labels are all locale keys), so no player met the error; only specs and outside callers could.
• The Titan Panel / broker display (GUI_TitanPanel.lua) is localized. Its stat and dataset labels, menu headers, broker label and tooltip lines were hard-coded English; they now read L[...], reusing the existing keys and adding nine (Recount Stats, Display Stat, Dataset, Overall, Showing, Left-Click, Right-Click, Toggle window, Configure display) to enUS, which every other locale falls back to. Found in Peer Review's single-pane read. Every flavour.
• The broker's Dataset menu now switches the data set the way the window's own fight menus do. It only set CurDataSet, so the main window and the detail window showed the old set until the next refresh, and the death log stayed on the old fight. It calls the new Recount:ShowDataSet(set, fightName) (GUI_Main.lua), which sets the fight name, updates the detail window, clears every pane (FullRefreshMainWindow), refreshes and points the death tracker at the new set. The six fight-menu entries in GUI_Main.lua still repeat that sequence inline and clear only the top pane's tables (the lower pane catches up on its data-set check); they were left alone. A split_spec example checks both panes empty and refill through it.

Development

• Language-server warnings cleared. .luarc.json declares ChatFrame1 and TooltipBackdropTemplateMixin, which the bundled LibDropDown reads. LibDropDown's options builder no longer assigns del()'s return value back to sortedOpts and lookup; it calls del and then sets each to nil. del already returned nil, so the behaviour is unchanged.
• wow-version-replication.ps1 no longer deletes a file from the other installs while it still exists here. An editor's save-by-replace removes the file for a moment; if the watcher's poll caught that gap it deleted every replica, then re-copied on the next poll. Sync-File now treats a "Deleted" whose source is back as a change. Reported by TOGProfessionMaster's review. The watcher was restarted and holds its lock; a deliberate save-by-replace was not reproduced.

Testing

• Eight weak assertions fixed: each checked that a key spelled as a literal was absent, which a typo would also pass. The key is now derived from the code, or the same key is also checked in a case where it has to be present. colors_spec compares the client's whole class-colour entry before and after GetColor; toc_spec refuses any field spelled like OptionalDeps, whatever its case; lazysync_spec also checks another player's version survives.
• toc_spec covers Recount_Camelot.toc, which it never read. Forever joins the per-flavour checks (interface 16001, same title, saved variables, no OptionalDeps), keeps Ace3 and VersionCheck-1.0 as its only dependencies, must load the retail file list plus the two bundled libraries, and is the only TOC allowed to load or declare them.
• The harness pin moved ff6e593 -> 870c83d, and Tests/env_recount.lua hands back everything it now provides (inbox thread 0fb44ef9, all six batches). Deleted: the unit readers (health, power, dead/ghost/feign, friend/attack, party/raid, threat), GetNetStats, the declension pair, GetInstanceInfo, GetZonePVPInfo, C_Scenario, C_PetBattles, IsEncounterInProgress, the combat-log readers, ReloadUI, CopyTable, GetMouseFocus, MouseIsOver, GetMouseFoci, GetSpellLink, the string and colour constants, the whole COMBATLOG_OBJECT_* / SCHOOL_MASK_* table, FauxScrollFrame_*, MinimapCluster, GetNumDisplayChannels, and our C_DamageMeter. Also gone, because nothing in Recount calls them: RegisterAddonMessagePrefix, the bare GetSpellName, GetMinimapShape (LibDBIcon falls back to "ROUND") and IsInScenarioGroup (deletion.lua checks for it). What stays is what the harness leaves out on purpose: UnitIsTrivial, YELLOW_FONT_COLOR, and the pre-Settings InterfaceOptionsFrame pair for the legacy-settings spec. Measured by resetting the harness modules without env_recount and checking each name on Classic Era and Mainline. The Classic-only globals (GetSpellLink, GetMouseFocus, MouseIsOver, SCHOOL_MASK_*) are now absent on Mainline, as on the client, and the suite stayed green.
  • env.setDamageMeterSession now writes the harness's wow.damageMeter model, and indexes each source by GUID and creature ID, since the harness keeps sources in their own table. A secret GUID is left out of that index. Our C_DamageMeter had been silently overwritten on every Mainline switch, which is what failed 36 retail and Forever examples at the new pin.
  • Two specs changed with the harness's answers: deletion_spec sets wow.zone, because the harness's GetInstanceInfo names the zone from it and an empty zone is Recount's "not loaded yet, retry" path; lazysync_spec stubs UnitIsConnected for its stubbed raid, which Recount asks by player name.
  • 816 passed, 0 failed. Recount.lua line coverage is 84.04%.
• modes_spec covers "Interrupted": the fixture adds a mob interrupting the healer, and two examples check the count, the interrupter and the interrupted cast, and that the mode's detail view reads InterruptedBy. The per-mode loop picks the new mode up on its own. 823 passed, 0 failed.
• Tests/tabs_spec.lua, 20 examples: the first tab mirrors the existing mode; + adds and selects; clicking a tab switches mode; other ways of changing mode relabel only the active tab; a tab keeps its mode when switching a module off shifts the indices; close, never-the-last, move with the selection; the right-click menu; the overflow row (< and > appear, hidden tabs, the selected tab in view, scrolling by arrow and wheel, arrows gone once wide enough); Show Tabs off; the short-name rules. Added after the self-audit and Peer Review: the tabs stay one level below the window when it is raised and never go negative; a row sized 4 units longer than the row but shorter than the window keeps every tab and the + inside it (red with either the inset fix or the trailing-gap fix reverted, measured); the Show Tabs checkbox lies inside its settings page; every mode label has an entry in the short-name table, on Classic Era and Mainline, so a new mode fails the suite instead of silently getting initials. 848 passed, 0 failed.
• Tests/split_spec.lua, first 17 examples: off by default; a data reset empties both panes; the lower pane shows the picked tab's display while the top follows the selected tab; the panes sit either side of the divider and inside the window; each pane's total bar; Show Below / Hide Below from the tab menu; closing, moving and closing-another-tab keep the lower tab right; a window too short for two panes drops the split and a taller one brings it back; dragging the divider with the harness's cursor grows the top pane and stops on release; a locked window does not drag; a click opens the detail window on the clicked pane's display (red with that routing reverted, measured); each pane's tooltip; a saved lower tab that no longer exists is repaired; retail Midnight draws both panes from the damage meter, and none below when the split is off. A menu entry that fades its menu shut raised offline, because UIFrameFade (every client's Blizzard_SharedXMLBase/FrameUtil.lua) was missing from the harness at 870c83d; a yielding stand-in covered it until the harness shipped the whole family.
• The harness pin moved 870c83d -> 27b84b3 -> 1867896 -> 6820bd5 and the UIFrameFade stand-in is deleted (inbox a4066a79). At 27b84b3 a frame with no height and a single anchor has no rect, as in the client, and GameTooltip had no height, so four tooltip-placement examples read nil; 1867896 sizes the tooltip from its lines (inbox f3ef297d) and they pass with no Recount-side workaround. 6820bd5 reports a spec file that hits the runner's time limit as a named TIMEOUT instead of crashing with (error object is not a string) (inbox 4fb957ea).
• The suite no longer keeps every earlier copy of Recount alive. env.initialize now drops Recount* globals whose frames came from an earlier reset (dropStaleNamedFrames), as a /reload empties _G. A named frame only some specs create, such as the compare window's RecountCompare_FightDropDown, kept its whole copy of Recount reachable through its initializer's upvalue (found by walking references from _G), and those copies went on receiving events. Across a full run the live frames grew from ~4,600 to ~47,000 and the heap to 140 MB, so modes_spec, run late, spent ~240 ms per example collecting garbage and passed the runner's 60-second limit. Now the frames stay at ~4,600 and the heap at ~17 MB throughout, modes_spec takes 13 s, and the whole run about a minute instead of about four. Clearing the names BEFORE the new load was tried first and broke 59 retail examples: a still-referenced old copy got an event mid-load and looked its scroll bar up by a name that was gone. The older "tooltip above the tab row" example, which contradicted the tooltip rule above, is removed with the operator's approval. 868 passed, 0 failed at the default limit.
  • env.initialize now fails with a named error when more than 12,000 frames are still live after the harness's reset (LIVE_FRAME_LIMIT), so a future leak fails the suite at its cause instead of timing out whichever spec runs late. Asked for by Peer Review.
• split_spec gains four examples (21 in all). One checks that Recount:ShowDataSet empties and refills both panes (see Fixed). Three come from Peer Review's reply: the divider drag stops when the button comes up off the divider, pressed and released through the harness's wow.mouseButtonsDown (the harness pin moved 6820bd5 -> b1f9386, which ships IsMouseButtonDown at Recount's request, inbox 7494ccfc; the stand-in env_recount carried until then is deleted); a fight fought after a data reset shows on the Current Fight view; and the lower pane follows both combat edges, showing the finished fight after combat ends and emptying when the next one starts. The combat-edge example first failed on its own setup: it fed the fight through TimeTick, and with nobody in combat by the test roster each tick ends the fight after one second of GetTime, which Recount discards as under three seconds (LeaveCombat -> CopyCurrentFights). The fight is now fed without ticks, and every assertion is unchanged.
• locale_spec checks every L["..."] literal in the files the Classic Era and Mainline TOCs load has an English entry. enUS is AceLocale's default locale, so it fills every gap in the other languages (the split view's three new strings show in English on a non-English client, per Peer Review's reading of AceLocale-3.0.lua:62-66 and :110-118); a key missing from enUS itself is the only one that would reach AceLocale's missing-entry path on every client. The scan finds over 500 reads and none is missing. It sees only literal keys: a computed read such as L[someVar], the shape of the GetModeIndex fix above, is invisible to it, and the spec says so.
• 873 passed, 0 failed across the whole suite at pin b1f9386.

[v1.19.1] (2026-09-27) - Forever login error

Fixed

• Forever: Recount errored on login, "zonefilters.lua:102: attempt to call a nil value". Reported from a Forever client. UpdateZoneGroupFilter called the bare GetZonePVPInfo, which Forever does not have; it ships only C_PvP.GetZonePVPInfo (wow-ui-source-forever, PvpInfoDocumentation.lua). The file-local now takes C_PvP.GetZonePVPInfo when the client has it and the bare global otherwise, so the older clients are unchanged. A spec covers the Forever shape, with the bare global made to raise.
• Forever: the other bare globals it lacks, found by comparing every global Recount's Forever file list reads against wow-ui-source-forever. Not yet run in a Forever client.
  • COMBATLOG_OBJECT_* exist on Forever only in Blizzard_DeprecatedCombatLog, which returns early unless the loadDeprecationFallbacks CVar is on. Recount.lua's FlagsToUnitID uses them as table keys while the file loads, so with the CVar off it would raise "table index is nil"; CreateOwnerFlags and AddCombatant add and mask them. Each now has an or literal like the flags above it.
  • Chat reports: SendChatMessage and BNSendWhisper are deprecation fallbacks on Forever too. GUI_Main.lua and GUI_Detail.lua now take C_ChatInfo.SendChatMessage / C_BattleNet.SendWhisper where the client has them (same arguments) and the bare globals otherwise.
  • The report window's RealID target called BNGetSelectedFriend / BNGetFriendInfo, which Forever does not have. The target is left off the list on a client without them.
  • The bundled LibDropDown called GetMouseFocus when the pointer left a submenu entry. It is gone from Forever and from retail 11.0+; the library now falls back to GetMouseFoci()[1].

Development

• wow-version-replication.ps1 replicates into _classic_beta_, the folder WoW Forever installs to (WowB.exe), so a Forever client picks up the working copy the same way the other flavours do. Verified with -DryRun, which lists it as a fourth target, and by the restarted watcher populating _classic_beta_\Interface\AddOns\Recount.

[v1.19.0] (2026-09-23) - World of Warcraft: Forever

New

• World of Warcraft: Forever (the _Camelot client, interface 16001). Recount_Camelot.toc loads the same file list as the Mainline TOC. Forever is the retail client -- WOW_PROJECT_MAINLINE, COMBAT_LOG_EVENT_UNFILTERED flagged HasRestrictions, and the same C_DamageMeter as Midnight (wow-ui-source-forever 1.60.1.69977, CombatLogDocumentation.lua, DamageMeterDocumentation.lua) -- so it takes the Midnight data path. The Midnight gate in Recount.lua, GUI_Main.lua, Tracker_Midnight.lua and GUI_Detail_Midnight.lua was interface >= 120000, which Forever's 16001 fails; it is now >= 120000 or < 100000, which leaves 11.x retail on the combat log. The test is made once, in Recount.lua, which publishes Recount.IsMidnightClient and Recount.IsRetailGame; the other four files read those instead of keeping their own copies, so they cannot drift apart. Without it Forever would have registered the restricted combat log at login, which is what threw ADDON_ACTION_FORBIDDEN on 12.0. Not yet run in a Forever client.

Fixed

• Forever: a click on a bar in combat would have errored. DamageMeterCombatSource.sourceGUID is ConditionalSecret on Forever (live leaves it unmarked), and the spell-detail window passed it straight back into C_DamageMeter.GetCombatSessionSourceFromType, whose SecretArguments = "AllowedWhenUntainted" refuses a secret from addon code. A secret GUID is now dropped before the call: an enemy is still found by creature ID, and a player with a hidden GUID shows the "available after combat" placeholder. The comments that called the GUID NeverSecret are corrected.
• Forever: the class-colour settings listed Death Knight, Monk, Demon Hunter and Evoker, because they keyed on the retail client rather than the retail game. Forever has the original nine classes.
• IsInScenarioGroup is called only if the client has it. Nothing on this box says whether Forever keeps it, so the retail branch of deletion.lua checks rather than assumes.
• Retail Midnight and Forever: the spell breakdown could error in combat on a hidden spell ID (audit finding 34). DamageMeterCombatSpell.spellID is not declared NeverSecret in the client's schema, and the getter is SecretWhenInCombat, but the breakdown used it as a table key unchecked -- the same "cannot be indexed with secret keys" error Recount users reported from the aura code. A secret spell ID is now skipped exactly like a secret amount (the window shows its "available after combat" state), and on the enemy side it is left out of the per-spell index while the enemy is still counted. The comment that called the field "never a secret" is corrected. A second comment, in the mitigation module, gave a reason for its isFullUpdate check that the schema rules out (finding 35); the check is unchanged and the reason now matches the schema.
• The bar-texture dropdown could be handed a texture name where it expects a list position. One variable held the saved texture's name and was overwritten with its list index only when found, so a saved texture no longer in the list (its media addon removed) reached UIDropDownMenu_SetSelectedID as a string. The two are separate now; a missing texture selects nothing. Every flavour.
• Reset Colors wrote into Blizzard's own class-colour table. ResetDefaultClassColors set a = 1 directly on the RAID_CLASS_COLORS entry (or a class-colour addon's CUSTOM_CLASS_COLORS entry) before storing it. It now stores a copy. Every flavour.

Testing

• Tests/forever_spec.lua: the gate on Forever and on 11.x retail, and the secret-GUID guard, with a spec double that refuses a secret GUID the way the client does. With the guard removed, the three guard examples fail.
• The harness pin moved 1f8fe09 -> e8a88cd (316 commits). One Recount-side break: installGlobals replaced _G.Enum wholesale, which deleted the harness's Enum.RegisterAddonMessagePrefixResult and failed 34 lazy-sync examples inside the harness's C_ChatInfo. It now extends Enum.
• The private test environment now overrides nothing the harness defines. Three harness defects found along the way -- IsInGroup ignoring its group category, and AbbreviateNumbers and FontString:SetText refusing a secret the client accepts -- were reported, wrapped locally for a few hours, fixed by the harness in 2c5f7fc, and the wrappers deleted. The pin is now ff6e593, which also delivers harness item 12 (text width scales with font size); no example moved. Handed back: the stubs it overwrote, the private bit library, C_Timer (whose stand-in only recorded calls and never ran one), ToggleDropDownMenu, the group, zone, guild and realm readers, RAID_CLASS_COLORS, SendChatMessage, the spell readers (specs now register the spells they cast, since the harness answers nil for an unknown spell as the client does), the per-flavour builds (Recount keeps one override: Mainline runs as 12.0), SlashCmdList, and the settings backend (the string-addressed stand-in survives only for the one example that models a client without C_SettingsUtil). env.time, env.locale, env.printed, env.chatMessages and env.openedCategories are views onto the harness's own records. What remains private is 59 globals the harness does not define yet (requested, and sequenced by WoWAPITesting). Deleting the colour stub exposed a spec asserting a fake MAGE colour (0.41; Classic Era's is 0.25).
• Secret values in the specs are real now. All 13 examples that used Recount's by-value registry (env.markSecret) use the harness's wow.secretValue, which raises on any arithmetic or comparison, and the registry is deleted. Under the old model a "secret" was an ordinary number that arithmetic still worked on, so a missing guard could not fail an example -- the gap round 8 of the audit named. The Midnight bar renderer, the spell-detail window and the mitigation module all pass with values that raise on contact. Two harness display sinks raised on a secret the client documents as accepting (AbbreviateNumbers and FontString:SetText, both AllowedWhenTainted); the harness fixed both in 2c5f7fc.
• Peer review and harness requests moved to writ's inbox. docs/AUDIT.md and Tests/HARNESS_CONTRACT.md were imported into the inbox and deleted (135a238); their full text stays in git history, committed first in 31e638d.

[v1.18.8] (2026-08-25) - the reset button actually resets on retail Midnight

Fixed

• Retail (Midnight): the Reset confirmation's Yes button did nothing. Reported from a live client and reproduced here: the reset icon opened "Reset Recount? Do you wish to reset the data?", and answering Yes left every bar exactly where it was. The reset routine clears Recount's own combatant tables -- and on 12.0 those tables are empty, because the client took the combat log away and the numbers on screen are read live out of Blizzard's own meter session instead. So the reset was wiping a store that held nothing and then re-drawing the window from the session it had never touched, which on screen is indistinguishable from the button being dead. The reset now also clears the game's combat sessions, through the same call Blizzard's own damage-meter window uses for its "reset all sessions" menu item, and it runs before the redraw so the window comes back empty rather than being repainted from the data we just failed to clear. Confirmed working in a retail client.

  Every Classic flavour and pre-12.0 retail are untouched: the new step is gated on Midnight, does not exist at all on the Classic builds, and reports failure quietly rather than erroring on a retail client whose meter API is absent.

Documentation

• The README and the CurseForge description were audited against the code, and both were wrong about reporting. Each claimed the report window offers "Say, Party, Raid, Guild, or a custom channel". The real list (GUI_Report.lua:34-50) is Say, Party, Instance, Raid, Guild, Officer, RealID, Whisper and Whisper Target, plus every chat channel you are joined to, with destinations you cannot currently use omitted from the list. Both now say so.

• Two shipped features were documented nowhere. Ctrl+click a bar pops out a live per-player graph -- it has been in the bar's right-click menu as "Show Realtime Graph (Ctrl Click)" the whole time and in neither document. And the live meter windows were one line under Reporting on the CurseForge page, which is also the wrong section; they now have their own, naming what they can actually track (per player, the whole raid's DPS/DTPS/HPS/HTPS, and FPS / latency / traffic / remaining addon bandwidth). Also added: how to drive the window at all -- the CurseForge page had no click, shift-click, right-click, alt-wheel or resize instructions anywhere -- along with Keep Only Boss Segments, the who/where recording filters, and the auto-clear-on-instance/group options.

• The Midnight explanation now has a permanent home instead of living in a patch note. The one full account of what does and does not work on 12.0 sat inside the v1.18.6 entry under Recent Updates, a section this repo holds to the last five patches -- so it was two releases from being deleted, taking the explanation with it. It is a standing section now, updated for what has changed since: the breakdown naming its targets, and reset clearing the game's own sessions.

Testing

• Five examples cover the reset path on the Midnight flavour, and the first one is the bug. It renders a session, asserts a bar is on screen, resets, and asserts the session is gone -- not the combatant table, which was always empty and is precisely why the old code looked correct. The rest drive the confirmation dialog's Yes button end to end (the actual reported entry point, which no example had exercised), assert the window is left with no bars, assert a client with no meter API reports failure instead of throwing, and assert the whole thing is absent on Classic. The offline environment gained the ResetAllCombatSessions half of C_DamageMeter, which it had never modelled.

[v1.18.7] (2026-08-25) - the retail 12.0 fixes: secret aura data, a removed mouse-over helper, an updated LibGraph, and a spell breakdown that finally names what you hit

Fixed

• Retail 12.0: Recount no longer throws every time an aura changes on you or anyone in your group. Reported as 27 errors in one session: TrackerModule_Mitigation.lua:198: attempt to perform boolean test on field 'isFullUpdate' (a secret boolean value, while execution tainted by 'Recount'). The 12.0 client can hand an addon aura data whose fields are secret -- readable enough to hand straight to the game's own display widgets, but illegal for an addon to test, compare or use as a lookup key. Damage Prevented does all three: it asks whether an update is a full refresh or a delta, and it files each aura under its spell id and its instance id. Each of those is now checked first and skipped when the client will not let us read it, so the tracking simply goes quiet on that client instead of erroring. isFullUpdate was only the first line reached -- the two ids sit two lines further down and would have thrown next, so all four reads are guarded rather than just the one in the report.

  What this costs, said plainly: on a client that makes the fields secret, Damage Prevented records nothing for the auras it cannot read. That is an undercount and never a wrong number, which is the same direction every other unknown in that mode is resolved. It affects retail only; no Classic flavour makes anything secret, and nothing about their behaviour changes.

• Retail 12.0: the graph windows no longer error continuously while the cursor is near them. MouseIsOver -- the function that answers "is the mouse over this?" -- was removed from the retail client in 12.0, and four places in Recount grabbed it once when the addon loaded and then called it forever after. Because two of them are per-frame handlers, that is not one error but one per frame for as long as the window is open; the same defect in our embedded copy of LibGraph produced 148 in a single sitting. All four now call the widget's own IsMouseOver method, which is what the removed function did internally anyway and which every supported flavour has had all along -- so this is one spelling on all six versions, with no version check and no fallback. Affects the damage-over-time graph, the comparison graph, and LibGraph's pie chart.

• Retail (Midnight): the spell breakdown no longer switches which spell it is showing you, without moving the highlight. Clicking a spell in the breakdown opens its per-target split below. The window remembered your choice as a ROW NUMBER, and the rows are ranked by damage -- so the moment another spell overtook the one you had selected, the refresh re-selected the same row, which was now a different spell. The highlighted row did not move, so nothing on screen said anything had changed: you would be reading one spell's name against another spell's targets. It now remembers the spell itself and follows it up or down the list.

• Retail (Midnight): the spell breakdown now names the mobs you hit, and how much you did to each. Clicking a spell showed a single target row with no name against 0 damage. The damage-done data genuinely has no target side -- the game hands addons one target slot per spell there and leaves it blank -- but the information does exist, in a separate part of the meter that lists each ENEMY and, beneath it, the spells that hit it and who cast them. Read backwards, that is exactly "which mobs did I hit, and for how much", and the lower list is built from it now. Three things had to be right for it to work, and each was wrong in turn: the entries on that side are keyed by a spell id that comes back as 0, so a per-spell lookup matched nothing and the list stayed empty; the per-target figure lives on the spell entry's own total rather than in the target record, which is blank there too, so the row first arrived correctly named and reading zero; and the percentage was still being taken against the selected spell's total, which is not a share of anything once the list spans every spell. The lower list now reconciles with the Damage Taken view of the same fight.

  Two limits worth stating. Matching a spell to you is done on your name, because that record carries no unique id -- so during a raid or dungeon encounter, where the game hides names from addons, the match cannot be made and the list falls back to what it showed before. And where the game reports no usable spell id, the list is everything you damaged rather than just what the selected spell hit; that is broader than the row above it, never a wrong name.

• Retail (Midnight): per-second figures below 1,000 no longer print all of their decimals. The breakdown's DPS column showed 161.82124328613 on one row and 5.2K on the row above it. The game's own number-shortening only shortens at 1,000 and above and hands anything smaller straight back, so a fractional rate arrived at the column with every digit intact. Rates are rounded to whole numbers before the column formats them.

• Retail (Midnight): the spell breakdown's column headings sat outside the window. "Name of Ability", "DPS", "Damage" and the two below them were drawn above the window's top border, across the title bar and the close button. The window is a port of the Classic detail window, and the port dropped the container frame that everything in the original hangs off -- a frame 32 pixels shorter than the window, bottom-aligned, so that its top edge starts below the title strip. Every position inside is measured from that container, so with it gone the whole contents -- headings, rows and both pie charts -- were laid out 32 pixels too high. The container is back, so the layout matches the Classic window it was copied from.

• Retail (Midnight): spells tied on exactly the same total no longer swap places on their own. With two spells on equal damage the ordering was left to chance, and this window re-sorts several times a second while it is open -- so tied rows could visibly trade places on a tick where nothing had happened at all. Ties now break consistently, so the list only changes when the numbers do. Ties are commonest at the start of a fight, when several spells have landed once each, which is exactly when the breakdown is most likely to be open.

Libraries

• The bundled LibGraph-2.0 is replaced with the shared copy, and it now supersedes every other copy on the client. Recount embeds LibGraph-2.0 for the damage-over-time graph, the comparison graph and the pie charts. The copy shipping here is the merged shared build, carrying the retail 12.0 IsMouseOver fix, and it registers in a +2000 version band.

  That band is the point. Only one LibGraph is ever live in a session: every copy registers under the same library name, the highest version number wins, and each loser stops at its own first few lines without defining anything -- silently. Measured across this install on 2026-08-25, the field was Details at 90062 and Recount at 90068 (both stock, both carrying the crash) and FastGuildInvite at 91068 (a private fork, also carrying it). FastGuildInvite's was therefore the copy actually driving Recount's graphs, which is why the earlier fix to Recount's own copy changed nothing in game. At 92068 this one wins outright, so installing this release fixes the graphs even for players who also run an addon shipping an older LibGraph. The bands are now +0 stock, +1000 a private per-addon fork, +2000 the shared copy.

  Only the .lua changed: all eleven textures were compared byte-for-byte against the shared copy and are identical. The library also gains two hooks Recount does not use yet -- a label hook for localising the axis numbers, and optional X-axis labels -- which are additive and change nothing about current behaviour. This is a stopgap while the library goes out as a proper external dependency; the version bands stop mattering once there is only one copy to install.

Testing

• The mitigation suite gained seven examples covering the two paths above. Six drive each secret field individually -- an unreadable full-update flag, a secret spell id, a secret instance id, a secret id during a full rescan, and a secret caster -- and each asserts the resulting state rather than merely that nothing threw, because the failure worth catching is a guard that skips the wrong thing. The seventh drives the module's own event frame, which is where the reported crash actually entered and which no example had exercised before: calling the tracking function directly cannot catch a dispatcher that stops forwarding the payload.

• Harness pin moved b307a9b -> 1f8fe09, and contract 22 was withdrawn because it was already delivered. The window suite reddened the moment the IsMouseOver fix landed -- the shared harness stubs the removed MouseIsOver global and, at our pin, did not stub the widget method it forwarded to. Contract 22 was raised for it; the harness had shipped exactly that the same day, on the Region type so textures and font strings get it too, in an adoption entry that cites this addon's crash. Our pin was six entries behind, which is why the request was written at all. Whole suite green after the move. The staged stand-in in Tests/env_recount.lua stays for now and is written to yield on its own: every Recount spec loads that private environment rather than the harness's frame layer, so the harness's method does not reach it -- removing it is part of the environment migration already on the repo's todo list, not a loose end from this release.

  Also picked up in that range: the harness now runs after_each when an example fails. Previously one red example could leave a substituted global installed for every later spec file, turning a single genuine failure into a wall of unrelated ones. Recount's counts did not change; nothing here depended on it.

• The Midnight suite gained eight examples, every one of them written from a defect seen in a live client rather than imagined. Both column headers are asserted to sit below the title strip (driven red by giving the container the window's full height, which reproduces the 32-pixel shift exactly); the enemy-side target lookup is covered by its real shape, including the spellID = 0 that made the first implementation match nothing, the amount that has to come from the spell entry rather than the target record, and a percentage that must be a share of the target list rather than of one spell; and a sub-1,000 rate is asserted to render as 162 rather than 161.82124328613.

  Two of them are honest about what they do not catch, which matters more than the count. The row-position example stays green under the mutation that reddens the header one -- the rows travel with the header, so their relative positions stay valid while the whole block is misplaced -- and that measurement is written at the call site so it is not mistaken for a second catch. Writing the aggregate example also found a real bug before the code shipped: the substitution was being applied inside the per-spell loop, so a second entry for the same spell appended into the shared aggregate table and corrupted the list for every other spell.

• A fixture that could only produce the happy shape is why none of this went red earlier. The spell builder in the Midnight suite took a target name as a parameter and every call passed a real one, so the blank record the live client actually sends was never exercised. The general form is worth keeping: when a fixture builder takes a value as a parameter, ask what the live source puts there when it has nothing to put.

• Reported the same LibGraph defect to FastGuildInvite, which embeds the same fork and produced the error the user saw, as finding S36 in its docs/AUDIT.md. Its copy is that addon's to fix.

• Answered peer-review round 7 (findings 31, 32 and 33), each driven red before being fixed. Both breakdown-window fixes above had their production change reverted and the suite re-run to prove the new examples catch them: the selection example failed with the wrong spell under an unmoved highlight, and the ordering example failed with raw encounter order. The pre-existing selection example stayed green under both mutations, which is the reviewer's process point demonstrated rather than accepted -- it is retitled for what it actually pins. Finding 33's remedy is a shared env.widgetsCreatedBy helper in Tests/env_recount.lua, so the two spec files that ask "did this click build a menu" can no longer answer it two different ways; one of them had been counting the whole widget registry, which is both unsound against a weak-valued table and a weaker claim than counting what the click produced. Suite 758 -> 771 passed.

[v1.18.6] (2026-08-19) - Mana Given, Damage Prevented, a live Midnight spell breakdown, and correct interface versions on every flavour

Added

• New "Mana Given" display mode: who gave the mana, not just who received it. Every power gain was recorded against the player who received it, so a question like "how much mana did my Judgement of Wisdom restore to the group this run" had no answer -- the paladin's own record held nothing, and the total was scattered across each recipient. Healing has always recorded both directions (that is what makes "Healing Done" and "Healing Taken" separate modes); power gains only ever had the receiving half. Now they have both. Bars in the new mode are the players who granted mana; clicking one breaks it down two ways -- by ability (Judgement of Wisdom, Vampiric Touch, Innervate, Mana Tide Totem) and by recipient. Mana only, deliberately: rage, energy and focus are self-generated, so a "given" view of them would never have a row in it.

  Note this records from this version onward. Fights already in your saved data have no giving-side numbers and will show blank in the new mode; nothing is lost or converted, it simply starts accumulating.

• New "Damage Prevented" display mode: how much damage a mitigation buff actually stopped. Flat and percentage damage reduction is invisible in the combat log -- the server takes it off before the event is written, so a swing that would have hit for 530 arrives as 500 and looks identical to one that was always 500. But it is still countable, because the reduction is a known constant: a melee swing that lands for anything at all while Stoneskin is up had exactly that constant prevented, and for a percentage buff the pre-reduction number follows from the one that landed. So the number is arithmetic, not an estimate. Bars are the players whose buff did the preventing; clicking one breaks it down by ability and by who it was prevented for. Armour is deliberately not counted: its reduction changes with attacker level and armour value, so it is not a constant and the same arithmetic does not hold.

  The buff list starts small on purpose. Every entry needs its exact per-rank value confirmed before it goes in, because a wrong number here produces a confident, plausible, wrong figure that nobody can spot by looking -- and a buff that is not listed is simply not counted, which is the right way round to be wrong. Shaman Stoneskin is in; more follow as each is verified. Like Mana Given, this records from this version onward.

Fixed

• Retail: moving the opacity slider on a Realtime window's colour dropdown no longer throws a Lua error. GUI_Realtime.lua captured the global OpacitySliderFrame when the file loaded and then used it without checking. That global does not exist in the retail client at all -- zero occurrences anywhere in its source -- so on a Midnight client the captured value was nil and the opacity handler errored the moment it ran. It was reachable in normal use: the guard on the neighbouring colour handler is skipped on exactly the paths that open a picker with opacity enabled.

• Window transparency is no longer applied backwards in one of the two places that read it. colors.lua read the picker's alpha as 1.0 - slider value on the pre-Dragonflight path, while GUI_Realtime.lua read the same slider straight. The client settles it: on Classic Era ColorPickerFrame:GetColorAlpha() returns the slider value unchanged, so GUI_Realtime.lua was right and colors.lua was inverted. Both files now call GetColorAlpha(), which exists on every flavour and hides the difference between Classic (the opacity slider) and retail (the picker's own colour widget). Nothing in Recount references OpacitySliderFrame any more, so the two files can no longer drift apart -- there is only one expression left.

• Power gains are no longer dropped because the other participant is untracked, in either direction. One shared routine records every power gain, and it is reached from two events that sit at opposite ends of it: an energize is somebody granting you mana, a drain is you taking it from somebody. The check that decided whether to record anything was a single fixed choice, so it was right for one of those and inverted for the other. Mana granted to you by an untracked caster used to vanish; correcting that alone would have started losing a warlock's own drained mana whenever mob tracking was off. Which participant to check is now decided by the event rather than assumed once, so both cases record. Drained mana is also no longer filed under "Mana Given" for the mob it was taken from, which would have credited a boss for giving away mana somebody took off it.

• The graph window's legend was broken on every flavour, and is fixed. One anchor point in GUI_Graph.lua was spelled "Right" instead of "RIGHT". Anchor points are a fixed set of names and the game rejects a mis-spelled one outright, so building the graph window raised an error at that line every single time. Recount catches errors during startup so the rest of the addon still loads -- which is why this showed up only as a small orange "init warning" in your chat frame at login rather than as a broken addon -- but the error stopped the graph window's construction three lines before it created its legend rows. Anything that then tried to read those rows failed too. If you have ever seen a Recount init warning mentioning the graph window, this was it.

• Shield and death records no longer vanish mid-fight either. The same sweep as below reached three more places: the record of who shielded whom, written when a shield goes up and when it is refreshed, and the death record itself. Shields are the worst case of the three, because that record is deliberately kept even for someone your filters exclude -- so that a tracked player's detail view can still show who shielded them -- which means it is the one record that could be accumulating and be swept away at the same time. A death is also, by its nature, likely to be the last thing a combatant does for a while, which is exactly the state the sweep collects. All now hold their place.

• Mana Given, Damage Prevented and absorb credit no longer vanish mid-fight when the person being credited does nothing else. Recount deletes a combatant it can find no other reason to keep after thirty seconds, and it keeps that clock ticking from a "last seen doing something" stamp that every recorder writes when it credits someone. Three did not: the new Mana Given side, the new Damage Prevented side, and shield absorbs. Somebody credited only by one of those -- a paladin whose Judgement of Wisdom is proccing mana off other people's swings, a shaman whose totem is soaking hits, a priest whose shield is eating damage -- looked idle to that sweep however much they were contributing, and their whole record could be dropped and then start again from zero part-way through a fight. All three now stamp. The comment claiming the omission was deliberate was wrong on the facts: that stamp is a liveness marker for cleanup and has nothing to do with the "active time" that per-second figures are divided by, so writing it cannot inflate anyone's DPS or HPS. This only ever bit players who had turned time/graph data on for a tracked type, because the sweep does not run otherwise.

• With "Merge Pets w/ Owners" on, a pet's damage now appears in its owner's breakdown instead of only in the bar total. The merge added the pet's total to the owner's bar and stopped there, so the bar read owner+pet while the ability list underneath it showed the owner's abilities only and added up to less -- with nothing on screen saying the difference was the pet. On a hunter that looks exactly like the pet not being recorded at all, which it always was: every pet has its own combatant record the whole time, and unticking the option shows it as its own bar. The breakdown now reconciles with the bar it sits under. Pet rows are listed separately and named, as Claw (Fluffy), rather than folded into the owner's own abilities -- a pet's Claw is not the hunter's ability, and merging the two would merge the per-hit detail behind each of them too. Applies to damage and healing, and to the "Damaged Who" / "Healed Who" views alongside the ability lists.

• Retail (Midnight): the spell breakdown window no longer sits there showing a finished fight's numbers. Clicking a bar on 12.0 opens a spell breakdown for that player, and nothing ever refreshed it. Every function that touches that window after it opens re-draws the rows captured at the instant of the click -- so it could be scrolled, re-sorted and clicked through all day and still be showing the same frozen snapshot. In practice: open it mid-fight, keep fighting, and the main window climbs while the breakdown stands still; the fight ends, the meter resets, the main window empties, and the breakdown is left listing the dead fight's spells with nothing on screen saying which of the two is live. On every other version the same sequence keeps both windows in step, because the ordinary refresh ends by re-filling the detail view and the Midnight path returns 200 lines before it. It now has that step of its own: the window re-reads live data on every meter update, empties when the meter is reset, and keeps the spell you had selected rather than jumping back to the top one each time. This is a retail-only path; no other flavour loads it.

• A Lua error on death is fixed. When someone died with no attributable killer and no killing ability, and Recount had already noted a second death event for them moments earlier (the pattern a Spirit of Redemption or a raised ghoul produces), the death handler compared a value that was never set against a number and threw. The comparison is now made from the recorded time of that earlier event, and a missing one means "not a double death" rather than an error.

• A memory leak that also pinned a per-frame handler on permanently is fixed. Recount keeps a short list of auras to ignore and sweeps it every fifth of a second, dropping entries older than ten seconds. The sweep was clearing a copy rather than the list itself, so nothing was ever actually dropped: the list grew for the whole session, and because it could never be seen to be empty, the repeating handler that does the sweeping never switched itself off either. Both stop now. Longer sessions were paying for this continuously.

• Recount is no longer flagged out of date on Classic Era, TBC/Anniversary, Wrath, Cata or retail. Four of the six manifests declared a stale ## Interface, and one of the four was the unsuffixed fallback that serves any flavour without a file of its own. Recount.toc moves 11508 -> 11509, Recount_Cata.toc 40400 -> 40402, Recount_Wrath.toc 30403 -> 30405, and Recount_Mainline.toc gains 120100 alongside its existing list. Note 11508 is the Vanilla Test interface and is lower than live Vanilla's 11509, so the base manifest was not merely a patch behind, it was pointing at a different client. Every value was re-derived two independent ways that agree: the installed clients' own build manifest (.build.info -- retail 12.1.0.69382, Classic Era 1.15.9.69109, Anniversary 2.5.6.69110) and the interface-version table on warcraft.wiki.gg for the three flavours with no install on this machine.

• Recount_BCC.toc is renamed Recount_TBC.toc (Interface 20505 -> 20506), which is the only spelling a client actually reads. The separator is part of the special filename: modern suffixes take an underscore and only the two legacy suffixes (-BCC, -WOTLKC) take a hyphen, so _BCC was neither form and was almost certainly never read by any client rather than having stopped working. The legacy -BCC spelling is dead in any case as of Classic Anniversary patch 2.5.5. With no recognised name, TBC and Anniversary clients fell through to Recount.toc and were handed a Vanilla manifest -- and because Anniversary is an installed, active product here, that fall-through was live rather than theoretical. _TBC covers TBC Classic and Classic Anniversary; no separate file is needed.

• wow-version-replication.ps1 no longer copies Tests/ and docs/ into the other WoW installs. Convert-GlobToRegex set its directory flag only from a trailing backslash -- which the BigWigs packager forbids writing in .pkgmeta, so .pkgmeta correctly uses bare names and the flag was never true. A bare Tests therefore compiled to ^Tests$, a pattern that matches only a file of that name, so the folder was listed in .pkgmeta, reported as loaded at startup, and replicated anyway. Bare non-wildcard entries are now resolved against the repo and promoted to directory patterns. The packaged CurseForge zip was never affected; .pkgmeta was correct all along and this was the second, drifted implementation of the same rules.

• The same script no longer copies dotfiles either. It skipped only the four git dotfiles by name, so .busted, .luacheckrc, .luarc.json, .markdownlint.json and .markdownlintignore were all being replicated, and any dotfile added later would have been too. It now prunes any path component beginning with a dot, matching what the packager does unconditionally -- a rule that cannot be expressed in .pkgmeta at all, since dot-prefixed entries there are silently ignored. This one is load-bearing rather than cosmetic: .git in these repos is a one-line pointer file to a git directory kept outside the WoW tree, and replicating it aims the copy at the wrong repository.

• Removed 495 files of previously replicated Tests/ and docs/ material from the _anniversary_, _classic_ and _retail_ installs, left there by earlier syncs before the two fixes above. Nothing the game loads was involved and the source install was untouched.

Testing

• The offline environment stopped carrying its own copies of five things the shared harness now provides. Tests/env_recount.lua had stand-ins for the aura API, number abbreviation, the per-unit-token model, group-channel addon messages and the colour picker, staged while the harness was asked for each. All five are now the harness's and the copies are deleted, which is the point of the migration: one implementation, so the thing the tests exercise is the thing every other addon exercises too.

• One of those copies had been quietly cancelling the fix it stood in for, and the way that was found is the useful part. Three predictions were made about which tests would break when the shared code arrived. Only one broke. That looked like good news and was not: two of the stand-ins replaced their globals unconditionally rather than yielding, so the new code never ran at all and the tests passed on the old copies. The tell for a stand-in that is masking a fix is a prediction that fails to fire, not a test that fails -- a green suite proves nothing until the replacement is known to be reachable. With them deleted, all 35 sync tests exercise the real message echo, including the one that checks Recount learns its own version from its own broadcast, which is the whole reason that behaviour was requested.

• Three specs pin the merged-pet fix listed under Fixed above. They assert the ability breakdown adds up to the bar it sits under, that the pet's rows are named, and that the breakdown is untouched when merging is off. Neutralising the fix reproduces the reported symptom exactly -- bar 700, breakdown 500, one row instead of two -- while the merging-off spec stays green, which is what shows the other two are pinning the merge rather than something both settings share. The pet tests that existed before counted bars (two, one, none) and could never have caught this, because the whole defect lives inside a single bar.

• Tests/toc_spec.lua now asserts the live interface number per flavour, not just a plausible range. The spec already checked interface versions and passed on all four stale manifests, because it only required the number to fall inside a 10000-wide band -- 20505 sits inside 20000-29999 exactly as comfortably as 20506. A range that wide can only catch a flavour mix-up, which is not the failure that occurs; the failure that occurs is a number one or two patches behind. Each flavour now carries the live value, asserted to be among those declared, with the measurement and its date recorded in the file so a later reader knows what it rests on. The range check is kept alongside it because the two fail differently and the difference is the diagnosis. Six new assertions.

• The same spec now checks each manifest against the repository, not just against the disk, because those are different questions and only the first was being asked. A test called "loads only files that exist" opens each path a TOC lists and passes if the file is there -- and every file is there, on the machine the work was done on. It says nothing about whether the file is in the repository, which is what the released zip is built from. That gap came within one commit of shipping: all six manifests were tracked, modified, and already listed TrackerModules\TrackerModule_Mitigation.lua, while the module itself had never been added to git. The package would have carried six manifests pointing at a file no clone has, on every flavour, and the suite would have stayed green throughout. Each flavour now also asserts that its own TOC is tracked and that every file it lists is tracked; if git cannot be consulted at all the test says so in those words rather than reporting a catastrophe. Made to fail first, and the proof is that the older test passed in the same run.

• The offline picker now models the client's alpha contract instead of a stored value, which is what made the inverted read survive. Tests/env_recount.lua answered GetColorAlpha() from a field it kept itself, on both picker shapes -- something no client does. A spec therefore could not observe the opacity slider at all, so the suite was free to assert the inverted convention indefinitely, and did. The Classic-shaped picker now returns the slider's value as the real client does; the retail-shaped one keeps the stored field as a stand-in for the picker's own colour widget. The spec that asserted 0.25 for a slider at 0.75 now asserts 0.75, and its comment records why.

• The offline suite's widget layer is now the shared harness's, not a private re-implementation, and that is what found the graph-window bug. Tests/env_recount.lua carried its own ~240-line widget model whose SetPoint simply stored whatever arguments it was handed and whose GetLeft/GetRight/GetTop/GetBottom returned made-up numbers. Nothing that mis-spelled an anchor point, or that depended on where a frame actually sat, could ever fail a test. The harness resolves real rectangles from the anchor chain and rejects an invalid anchor exactly as the client does. Swapping the two took the suite from 712 passing to 80, and every one of the 632 failures was traced to a cause: one shipped defect, two harness gaps, and a set of specs that had been asserting against the private model's fictions. All 712 pass again.

• Several specs were reading methods the game does not have on the widget they were reading them from. The private model served one flat method table to every widget type, so a StatusBar answered GetVertexColor and a Texture answered GetText -- neither of which is true in game. Those specs now read the API the client actually exposes (GetStatusBarColor for a bar), and two of them were verifying that a value had landed somewhere that does not exist.

• Five further specs cover the liveness stamp on the absorb, shield, death and damage-prevention paths, again all made to fail on purpose first. Each asserts the stamp is current rather than merely present, which matters because every fixture registers its combatant with an event that already stamps it -- an assertion against "not nil" would have passed with the fix removed. Each of the seven live stamps was then disabled on its own and the whole suite re-run, seven times: every one of them takes at least one spec red on its own, so not a single one is unverified. An eighth apparent call site, in the shield-guessing branch of the absorb path, turns out to sit inside a block that has been commented out since long before this release -- it is not a stamp at all, and an earlier note describing it as an untested gap was describing code that cannot run. The comment there now says so.

• Two new specs cover the liveness stamp on the giving side of a power gain, and both were made to fail on purpose before being trusted. One asserts the stamp directly; the other drives the idle sweep across a minute of ten-second gifts and asserts the running total. With the stamp removed the second reports expected: 60, actual: 10 -- the 10 being the gift that re-created the combatant after the sweep had deleted it, which is the failure in the shape a player would actually see it. The sweep's own preconditions are recorded in the spec: it only runs when time data is on for a tracked type, and the filter clause of its idler test cannot fire for a combatant that has recorded anything, because AddAmount gates on the same table.

• A spec of mine passed for the wrong reason and is fixed. One of the four new power-gain specs asserted is_nil against a combatant name that never existed in the fixture, so it verified nothing while reading as a passing test -- in a spec written to catch a real bug. It now asserts that the receiving side still records while the giving side does not.

• The linter is trustworthy again. .luacheckrc imported the language server's globals list wholesale into its read-only set, and that list contains _G. Since every file in the addon opens with local Recount = _G.Recount, luacheck read the addon's basic declaration idiom as writing to a read-only global -- roughly fourteen warnings per file across fifty files, a permanently non-empty report in which no real finding could be noticed. The import now filters out anything the addon legitimately writes. Several pieces of genuinely dead code turned up once the noise cleared and were removed -- and so did the death-handler error and the leak listed under Fixed above, which is where they belong; a crash filed as tidying is a crash nobody reads about.

• The unused-argument exemption is gone entirely, rather than being narrowed. Combat-log handlers are dispatched by position and must declare the client's full argument list whether or not they read each name, so the checker's unused-argument rule had been switched off for the whole addon. That was wrong twice over: the same switch also silences unused loop variables, which have no such excuse (measured, it was hiding 54 of those in shipped code, none in a handler), and switching a check off is not how to express a calling convention anyway. Every unread position now carries a leading underscore instead -- the name still records what the client puts there, the position is untouched, and the checker treats the omission as deliberate. 191 parameters across the dispatcher and the tracker modules, with no configuration exemptions left anywhere in the repo.

• Two blanket luacheck: ignore blocks in the dispatcher were replaced with actual fixes. One hid fourteen combat-log flag constants that were simply never used; they are deleted, and the copied-from-upstream table is now trimmed to what the addon reads. The other hid two near-identical copies of the absorbed-damage handler, one per payload shape; they are now a single body that reads its arguments by index, with the two layouts written out once in a comment above it. The suite covers both shapes and still passes.

• .markdownlint.json stops disabling the rules that decide how the release notes render. This file is published verbatim as the GitHub release body, so a blank line inside a quote silently splitting it in two, or a list running into the paragraph above it, is a defect in the product rather than a style preference. Four of the five disabled rules were of that kind and are back on; duplicate-heading checking returns in its standard changelog form, where a repeated ### Fixed under different versions stays legal. Only the line-length rule stays off, because prose that explains why runs long. Re-enabling them turned up three real violations repo-wide, all now fixed.

• The offline picker no longer defines a global the retail client does not have. Tests/env_recount.lua installed OpacitySliderFrame on both picker shapes, so the retail-shaped environment modelled a client that has never existed -- and that is the specific unfaithfulness that let the crash above ship, because a load-time capture of the global found an object offline and nil in the real game. The retail shape now omits it, which makes the existing spec a genuine reproduction rather than a guard against re-introducing the call-time form only.

• The offline environment was running every spec as Classic Era, whatever flavour the spec asked for. Tests/env_recount.lua selected the flavour and then called the shared harness's reset, which ends by restoring the flavour to Classic Era -- so the selection was overwritten before anything read it, and the line that looked like it restored the value was reading the harness's default back. Every version-dependent piece of the environment was therefore built for the wrong client on every run. Found by instrumenting the one call site that visibly disagreed and printing the value: a load asking for retail reported Classic Era. The requested flavour is now handed over directly instead of through a global the harness owns. The remaining half -- that the harness picks its own per-flavour widgets before a consumer can say which flavour it wants -- is raised as a harness contract item.

• The colour picker's shape now follows the client version, and two selectors became one. Which of the two colour-picker shapes a spec got was chosen by a bespoke option while the client version was chosen by another, so a run nominally on Classic Era was handed the retail picker -- and the one global that tells the two apart was therefore missing on the only version that has it. Two tests named for "a client with no opacity slider at all" were asserting retail's behaviour on a Classic Era run and passing. The option is deleted rather than kept as an override, because the only thing it could still express is a client that has never shipped; four call sites now name the version they mean.

• Fifty-six tests asserted nothing, and nine more had their assertion behind a condition that was never true. Each display mode was driven through four tests, two of which had no assertion at all -- "opens the detail window" clicked a bar and checked nothing -- while a third was wrapped in "if there are any bars", which for three modes there never are. The unfed modes were measured rather than guessed (Absorbs, Threat and Damage Prevented, the last two of which no reading of the fixture would have shown) and the list is now asserted in both directions, so a mode that starts or stops producing bars fails loudly. The rest assert what their names promise: that the click selects the combatant and puts the detail window on screen, and that the right-click builds a menu. 112 mode tests became 150.

• Ten more tests asserted that a window the setup had just opened was still open. Locking a detail row, selecting one, selecting a pie slice, walking every detail view, rendering each saved fight, plotting a graph series, toggling the compare window's options -- each ended by checking a window was shown, which nothing in the test could have changed. Three were worse than that: the helper that "opened" the detail window bypassed the code that fills it, so the three selection calls under test hit an empty-table guard and returned immediately, which is why only that assertion could pass. The helper now clicks the bar the way a player does, and each test asserts the state its subject actually writes. Two were then confirmed by deleting the code they cover and watching exactly one test go red each.

• Tests/mitigation_spec.lua is new, covering the Damage Prevented module: that credit lands on the buff's caster rather than the player who was hit, that a buff refresh arriving as a removal and an addition in the same event never leaves a gap the swings in between fall through, that a swing reduced to nothing is skipped rather than guessed at, and that a swing whose absorbed portion is filed as a second row is still only counted once. Twenty-one assertions.

• The offline harness pin moves 239 commits, from c46cb89 (2026-08-02) to ff379c2 (2026-08-16). The old checkout predated the harness's Ace3 loader, widget layer, library loader, coverage tool and verification tooling.

• Tests/coverage.lua is deleted; coverage now runs on the harness's own multi-target implementation. The local file was the staged reference implementation for a request the harness delivered on 2026-08-07, and an addon-local copy of a shared tool is how divergent copies accumulate across the suite. lua Tests/run_coverage.lua is unchanged as the entry point, but it is now only a driver: it enumerates the shipped files and hands them to Tests/wowapi/coverage.lua. The detail flag is inverted from the old copy -- uncovered lines print by default and COVERAGE_SUMMARY=1 suppresses them.

• .busted no longer tells the reader to run busted. The file itself is kept, because it is inert and is what stops an accidental invocation collecting recursively, but its header said to run the executable from the Recount root and that instruction gets followed. Doing so collects the harness's own self-tests and runs them from the wrong working directory, producing failures the addon did not cause. The one supported command is lua Tests/wowapi/run.lua.

• The two new display modes are now tested by value, not just driven. The existing mode sweep walks every registered mode through select, refresh, hover and click, which proves nothing raises -- and cannot see a mode reading the wrong field, because it still renders a title, still fills a tooltip and still opens a detail window. That is the likeliest defect in both new modes: Mana Given mirrors Mana Gained and reads four sibling keys, and Damage Prevented reads three keys written into the same table. Fifteen specs now assert the values that reach the bar, the two detail views and their order, the totals agreeing between the bar and the detail form, and the tooltip naming the right ability and the right player. Verified by mutation rather than by writing them green: swapping Mana Given onto the gaining fields fails four of them while all 112 generic mode tests still pass, which is the argument for having them. One of the fifteen exists only because that exercise found it -- the detail form's own total was asserted by nothing.

• A limitation the description page had stopped mentioning is back, in the right terms. The page carried a Midnight warning that had become obsolete -- it said the bars stay empty, which they no longer do -- and removing it left the page silent about what the client genuinely restricts. It now says plainly that on Retail Midnight the game hides combat numbers from addons, so Recount can display live per-player bars and open a breakdown on click, but cannot sort them itself, show a Total bar or percentages, or put a tooltip on a bar. Stated as the game's restriction, which is what it is, and scoped to Midnight so no Classic player reads it as applying to them.

Suite: 727 passing, 0 failing.

[v1.18.5] (2026-08-03) - MoP Classic support, TOC metadata corrections

Fixed

• MoP Classic gets its own TOC (Recount_Mists.toc, Interface 50504/50503). Recount branches on WOW_PANDA_CLASSIC in several places and both the README and the CurseForge description claimed MoP support, but there was no _Mists TOC -- so the client fell back to the base Recount.toc (Interface 11508), flagged the addon as out of date, and it only loaded at all with "Load out of date AddOns" ticked. Worse, the packager derives the game versions it publishes from the TOC files, so the v1.18.4 upload registered 1.15.8, 2.5.5, 3.4.3, 4.4.0, 11.2.7, 12.0.5 and 12.0.7 but no 5.5.x -- CurseForge never offered the build to MoP players at all. The new TOC is byte-identical to the Cata one apart from the interface version.

Changed

• Retail now requires VersionCheck-1.0, like every other flavour. Recount_Mainline.toc declared Dependencies: Ace3 while every Classic TOC declared Ace3, VersionCheck-1.0 -- and had never listed it in the file's history, so retail players got no in-game notification when a new Recount shipped. Both are hard dependencies now on all six flavours. CurseForge already installs both automatically (they are listed under required-dependencies in .pkgmeta), so this changes nothing for anyone installing through the app; a manual installer needs VersionCheck-1.0 present, as Classic players already did. Tests/toc_spec.lua asserts the dependency lines stay in step from here on.

• The bundled libraries are declared as embedded rather than as optional dependencies. Every TOC listed LibDropdown-1.0, LibSharedMedia-3.0, LibBossIDs-1.0, LibGraph-2.0 (and Classic Era alone also listed LibDataBroker-1.1, LibDBIcon-1.0) under ## OptionalDeps: -- but all six of those ship inside libs/ and are loaded by embeds.xml or by explicit TOC lines. OptionalDeps only reorders standalone addons of the same name, so those entries were inert and described the addon incorrectly. Replaced across all six TOCs with ## X-Embeds: LibStub, LibDropdown-1.0, LibBossIDs-1.0, LibSharedMedia-3.0, LibGraph-2.0, LibDataBroker-1.1, LibDBIcon-1.0 -- the field that actually means "bundled inside this addon", now identical on every flavour and naming all seven (the old line missed LibStub entirely and used the wrong capitalisation for LibDropdown-1.0). Tests/toc_spec.lua cross-checks the declaration against the contents of libs/, so a library cannot be bundled in future without being recorded.

Testing

• Tests/toc_spec.lua guards the flavour TOCs against drift. They are near-identical by design, so the failure mode is one TOC gaining a file the others do not (that flavour silently loses a feature) or a flavour having no TOC at all. The spec asserts every supported flavour has a TOC, that its interface version is in that flavour's range, that every file it lists exists on disk, that the Classic flavours load exactly the same files in the same order, that retail loads that same list plus the two Midnight files and no Classic flavour loads them, that all 15 locale files are listed, and that dependencies, saved variables, title and project id match across flavours.

[v1.18.4] (2026-08-03) - Addon Language selector, nine bug fixes, offline test suite

Added

• "Addon Language" dropdown in Settings. Recount's own interface text can now be forced to any bundled translation, independent of your WoW client language -- a player on an English client can run Recount in German, a German client in English, and so on. The dropdown lists Auto (the previous behavior -- follow the client locale) plus all 14 bundled languages: English, Deutsch, Español, Español (AL), Français, Italiano, Nederlands, Português (BR), Русский, 한국어, 简体中文, 繁體中文, Thai, Filipino. The choice is saved (RecountLanguage) and applies on the next /reload. In-game item / spell / unit names still come from your client and are unaffected.

• README.md now ships with the addon. A player-facing guide sitting next to CHANGELOG.md in the installed folder and on the GitHub page: what Recount does, what it needs installed alongside it (Ace3 and VersionCheck-1.0), how to drive the window, every display mode, the detail / graph / compare / death-log views, fight history, live meter windows, reporting, the minimap button and broker feed, the settings worth knowing about, the full slash-command and keybinding surface, the retail Midnight differences, and a short troubleshooting section. It is deliberately not in .pkgmeta's ignore list — it is meant to be read after installing, not only on the web.

Fixed

• Right-clicking the minimap button no longer throws an error when opening Recount's settings. On clients that ship C_SettingsUtil.OpenSettingsPanel (TBC and other flavors on the newer settings backend), Settings.OpenToCategory forwards its argument to C_SettingsUtil.OpenSettingsPanel, which requires a numeric category ID. GUI_Minimap.lua passed the literal string "Recount", producing bad argument #1 to 'OpenSettingsPanel' (outside of expected range …) from Blizzard_Settings.lua:144. Recount now uses the real category identifier handed back by AceConfigDialog when the panel is registered.

• The main window is no longer hidden when "Hide window when not collecting" is off. Recount:SetZoneGroupFilter — the combined zone+group filter that every zone change actually goes through — was missing a pair of parentheses: HideCollect and IsShown() or not GlobalDataCollect binds as (HideCollect and IsShown()) or (not GlobalDataCollect), so turning global data collection off hid the window even though auto-hide was never enabled. The zone-only and group-only filters had the correct grouping; all three now agree.

• Lowering "Max Fights" now trims the fight list even when there is no data. Fights:DeleteOverflowFights cleared Recount.db2.FoughtWho inside its per-combatant loop, so with zero combatants recorded (right after a reset, for instance) the fight names survived and the fight dropdown kept offering entries with nothing behind them. The name list and the death log are per-meter, not per-combatant, and are now trimmed once.

• Recount no longer writes into Blizzard's class-colour table. Colors:GetColor("Class", …) stamped a = 1 onto the RAID_CLASS_COLORS / CUSTOM_CLASS_COLORS entry it was reading, mutating a table shared with the default UI and every other addon. It now returns a copy.

• Clicking a bar on retail Midnight no longer throws. Recount:CreateMidnightDetailWindow never registered its window with Recount:AddWindow, so the frame had no Above / Below links; its own OnShow calls Recount:SetWindowTop, which then indexed window.Below — nil. Every attempt to open the spell breakdown from a bar errored. The window now joins the stacking chain like every other Recount window, and SetWindowTop falls back to UIParent's frame level when a window has nothing below it instead of erroring.

• Other players' pets no longer resolve to your own pet. Recount:FindPetUnitFromFlags tested affiliation with bit_band(COMBATLOG_OBJECT_AFFILIATION_MINE) — no mask argument. LuaBitOp returns the value unchanged for a one-argument band, so the test was MINE ~= 0, i.e. always true, and every pet resolved to the "pet" unit token. Any unit lookup for a raid or party pet (tooltips, health, threat) pointed at the player's own pet instead. Now masked against unitFlags.

• Disabling the addon no longer errors. Recount:OnDisable called Recount:UnregisterAllEvents(), but Recount embeds AceConsole / AceComm / AceTimer and not AceEvent, so the addon object has no such method. The client events live on the Recount.events frame, which is what is unregistered now.

• Recount:AddGraphNameEntry works. The public registration point tracker modules use to label their own graph series called GraphName.insert(k, v) on a plain lookup table, which has no insert field — every call errored. It assigns GraphName[k] = v.

• Recount:FlagSync no longer errors on a group member with no recorded data. It indexed the combatant record without checking it exists.

• A pet's owner is now recorded on the pet. Recount:SetOwner registered the pet on its owner but never set Owner on the pet itself — the field was only ever written as false (the "not a pet" marker), so three separate features that read it were dead: pet damage and healing were left out of the raid data sync entirely, deleting a pet left its name behind on the owner's pet list, and the "Merge Pets" option's owner inheritance never applied, so a pet stayed on the meter even when its owner's category was filtered out.

Testing

• Offline unit-test suite (Tests/). Recount now carries a busted-compatible spec suite built on the shared WoWAPITesting harness (added as a submodule at Tests/wowapi), runnable with nothing but a Lua 5.1 interpreter: lua Tests/wowapi/run.lua. 596 specs, all passing. The suite loads Recount's real shipped files — and the real Ace3 / LibDBIcon / LibGraph / LibDropdown libraries from the sibling AddOns folders — into an offline environment, then drives the addon through its actual entry points (combat-log events, minimap clicks, widget scripts, colour-picker callbacks, addon-to-addon sync messages) rather than testing reimplementations.
  • Tests/env_recount.lua stages what the shared harness does not yet provide: a bit library faithful to LuaBitOp (bare Lua 5.1 has none), a frame system whose unknown keys resolve to nil rather than to no-op functions, named children for FauxScrollFrameTemplate / OptionsSliderTemplate / GameTooltipTemplate, a client-style debugstack (LibGraph-2.0 parses its own stack trace to find its textures and errors without it), a WoW-compatible xpcall (stock Lua 5.1 drops the extra arguments ChatThrottleLib passes through it, so no addon message ever reached the wire), C_DamageMeter with 12.0 secret-value semantics, a UIDropDownMenu that actually runs its initializer, and selectable settings-backend / colour-picker contracts. Each is written up in Tests/HARNESS_CONTRACT.md for upstreaming; the submodule was not modified.
  • Tests/coverage.lua measures exact line coverage from Lua 5.1 bytecode debug info rather than a source-text heuristic. lua Tests/run_coverage.lua reports 16980/19275 executable lines (88.1%) across every shipped file; excluding the locale data files that is ≈83% of the code. Every file is above 63%, and 24 of 32 are above 84%: WindowOrder 98%, GUI_DeathGraph 98%, GUI_Detail_Midnight 97%, GUI_Reset 96%, Fonts 96%, PowerGains 96%, GUI_TitanPanel 95%, Resurrection 95%, Tracker_Midnight 94%, CCBreakers 94%, colors 91%, Fights 91%, zonefilters 91%, Interrupts 91%, Dispels 91%, roster 90%, deletion 90%, GUI_Config 90%, Threat 89%, Widgets 87%, GUI_Minimap 86%, Recount_Modes 84%, Recount.lua 84%, GUI_Report 82%, GUI_Graph 80%, Tracker 79%, LazySync 77%, GUI_Detail 77%, GUI_CompareGraph 74%, GUI_Main 73%, GUI_Realtime 69%. What is left is drag-to-move, drag-to-resize and code that branches on anchor-resolved geometry or real font metrics — all three need a layout/interaction model in the shared harness, written up as items 11–13 of Tests/HARNESS_CONTRACT.md.
  • Specs: smoke (loads and initializes cleanly on all six flavours, with no init warnings), core (lifecycle, events, slash commands, table pool, combat state), minimap, colors, windoworder, zonefilters, fights, roster, deletion, locale, tracker and combatlog (the combat-log pipeline and the whole dispatch table), modes (every display mode select → render → hover → click), mainwindow, interactions (dropdowns, formatting, detail tables, threat polling), guiwindows (graph, compare, death graph, reset, report, realtime, Titan Panel), config, midnight (12.0 secret values), lazysync (the sync protocol, asserted on the wire) and guidepth (graph series maths — integrate, normalize, stack, thin — plus compare-graph series and realtime window pooling).
  • Tests is excluded from the packaged zip.

Packaging

• Fixed .pkgmeta ignore patterns that were matching nothing. The BigWigs packager matches with shell case globbing, which has no recursive ** support, so "**/*.ps1" and "**/*.bat" never matched and wow-version-replication.ps1 was shipping to players. They are now "*.ps1" / "*.bat". The dot-prefixed entries (.git, .github, .vscode, .claude, .luarc.json, …) were no-ops — the packager prunes dotfiles unconditionally — and have been removed rather than left implying coverage they never provided.

Implementation

• Recount.lua captures the Blizzard category identifier. AceConfigDialog:AddToBlizOptions("Recount Blizz", "Recount") returns (frame, categoryID); the second value is stored as Recount.BlizOptionsCategory. This is version-agnostic by construction: on older AceConfigDialog builds (and clients without C_SettingsUtil) the library deliberately overrides category.ID to the category name, so the captured value is the string "Recount" and behaviour is unchanged; on newer builds the override is skipped and the captured value is Blizzard's generated numeric ID. GUI_Minimap.lua passes Recount.BlizOptionsCategory or "Recount" to both Settings.OpenToCategory and the legacy InterfaceOptionsFrame_OpenToCategory path, so the fallback still works if the panel registration ever fails.

• Every locale file now registers through ns.NewLocale (locales/Recount-LocaleCore.lua). AceLocale-3.0 keeps only the locale table matching GetLocale() (plus the enUS default) and discards the rest via if not L then return end, so a UI-language override had no strings to switch to. ns.NewLocale(code, isDefault, silent) records every L[key] = value into both ns.Locales[code] (always — so the override path has every language in memory) and AceLocale's table (only when it matches the client / is default — so auto-detect is unchanged). The 9 community locales (deDE, esES, esMX, frFR, ptBR, ruRU, koKR, zhCN, zhTW) were converted from the stock LibStub("AceLocale-3.0"):NewLocale(...) + if not L then return end header to local _, ns = ... / local L = ns.NewLocale("<code>"). (itIT, nlNL, thTH, filPH already used it.)

• English is selectable too. Recount-enUS.lua now routes through ns.NewLocale("enUS", true, debug). enUS uses AceLocale's L[key] = true convention (the value is the key); for that case ns.NewLocale stores the key string itself in ns.Locales (AceLocale still receives true), so forcing English on a non-English client overlays real English text rather than the boolean true. The existing debug flag is forwarded as AceLocale's silent argument.

• Load-time overlay (top of Recount.lua). On load, if RecountLanguage is set and not "auto", it copies ns.Locales[override] onto the shared L table before any other file captures L[...] into its own structures (mode list, option tables). RecountLanguage is a plain SavedVariable, readable at load before those captures happen. Keys missing from the chosen language keep the client/English fallback.

• Dropdown labels. Native-script names for Latin scripts (Deutsch, Español, Français…); CJK and Cyrillic carry an English hint in parentheses (한국어 (Korean), 简体中文 (Chinese, CN), 繁體中文 (Chinese, TW), Русский (Russian)) because WoW's default Western fonts render those scripts as boxes on a non-native client; Thai is shown as "Thai" for the same reason.

Notes

• The community translations (deDE, esES, esMX, frFR, ptBR, ruRU, koKR, zhCN, zhTW) are older WoWAce exports and are likely partial — untranslated and newer strings (Compare Graph, Midnight, etc.) fall back to English. They are now selectable, not necessarily complete.
• Every translation now stays resident in memory rather than only the client's locale — a small, intentional cost for the override feature.

Multi-version safety

• Pure locale-layer + options change; no version-specific code paths. Identical on Classic Era, TBC, Wrath, Cata, MoP Classic, and retail (incl. Midnight). With the dropdown left on Auto, auto-detect is byte-for-byte unchanged.
• The minimap settings fix carries no client check of its own — it reads whatever identifier the installed AceConfigDialog registered, so it is correct on every flavor regardless of which settings backend that client runs.

Files Changed

• locales/Recount-LocaleCore.lua — ns.NewLocale gains silent forwarding and key-as-value handling for the = true convention
• locales/Recount-enUS.lua — routed through ns.NewLocale("enUS", true, debug)
• locales/Recount-{deDE,esES,esMX,frFR,ptBR,ruRU,koKR,zhCN,zhTW}.lua — converted to the ns.NewLocale header
• Recount.lua — "Addon Language" dropdown expanded from 5 entries to all 15 (Auto + 14 languages); captures Recount.BlizOptionsCategory from AddToBlizOptions
• GUI_Minimap.lua — right-click opens the settings panel by category identifier instead of the hard-coded name
• zonefilters.lua — parenthesised the auto-hide condition in SetZoneGroupFilter
• Fights.lua — DeleteOverflowFights trims FoughtWho and the death log once, outside the per-combatant loop
• colors.lua — GetColor returns a copy of the client class colour instead of mutating it
• WindowOrder.lua — removed a dead v:GetScript("OnMouseUp") statement in ScaleWindows that fetched a handler and discarded it; SetWindowTop no longer indexes a nil Below
• GUI_Detail_Midnight.lua — the Midnight detail window registers with AddWindow
• Recount.lua — masked the affiliation test in FindPetUnitFromFlags; OnDisable unregisters events on Recount.events
• GUI_Main.lua — AddGraphNameEntry assigns into the label table instead of calling a non-existent insert
• LazySync.lua — FlagSync skips group members with no recorded data
• Recount.lua — SetOwner records the owner on the pet, not just the pet on the owner
• Tests/ — new offline suite (see Testing above)
• .pkgmeta, .luacheckrc, .busted, .luarc.json — packaging and lint configuration for the suite

[v1.18.3] (2026-06-10) - Retail Midnight (12.0) live damage meter display

Fixed

• Recount bars now populate live during combat on retail Midnight (12.0+). This completes the work scaffolded in v1.18.1, where the addon loaded cleanly but its bars stayed empty because C_DamageMeter returns secret combat values. The v1.18.1/v1.18.2 approach — poll C_DamageMeter.GetCombatSessionSourceFromType, diff against a snapshot, feed deltas through AddAmount — could never work: it did arithmetic (total - snapshot) and comparison (delta > 0) on secret values, which is exactly what taints addon execution, so it skipped every secret and applied no data. That whole snapshot/diff path is removed.

• New approach — display passthrough, no arithmetic. The rule (verified in-game): an addon may pass a secret value straight into a display sink — FontString:SetText, StatusBar:SetValue / SetMinMaxValues, AbbreviateNumbers — and it renders live, mid-fight, with no error and no taint. What an addon must never do is operate on a secret (arithmetic, comparison, sort, percentage, total, string op) or route a secret into a shared frame (above all GameTooltip); either silently taints the running execution, and that taint then poisons whatever shared resource it next touches (the classic symptom is an unrelated Blizzard tooltip — e.g. world-map POIs — throwing "Secret values are only allowed during untainted execution"). pcall does not help and is never used to "guard" a secret: it catches the throw but leaves the taint in place, which is worse because it's silent.

Implementation

• Tracker_Midnight.lua rewritten as a secret-safe renderer. On the DAMAGE_METER_CURRENT_SESSION_UPDATED / DAMAGE_METER_COMBAT_SESSION_UPDATED / DAMAGE_METER_RESET events (and via the periodic refresh, see below) it calls C_DamageMeter.GetCombatSessionFromType(Enum.DamageMeterSessionType.Current, metric) for the on-screen mode's metric and writes each combatSources[i] straight into the existing Recount.MainWindow.Rows[i] widgets: LeftText:SetText(src.name), RightText:SetText(AbbreviateNumbers(amount)), StatusBar:SetMinMaxValues(0, session.maxAmount) + SetValue(src.totalAmount). No value is ever computed on.

  • Uses the session API GetCombatSessionFromType (returns the whole combatSources array, pre-ranked) rather than the per-source GetCombatSessionSourceFromType the v1.18.1 scaffolding used.
  • Field secrecy honoured per the API docs: name / totalAmount / amountPerSecond and session maxAmount are secret in combat (passed through, never computed on); classFilename, specIconID, classification, isLocalPlayer, deathRecapID, sourceGUID are NeverSecret (used directly — class colour comes from classFilename).
  • Bar order is the API's pre-ranked order (we can't sort by a secret value). No Total bar and no percentages on Midnight (both need addition). #combatSources is a plain count, safe for scrollbar maths and row indexing.
  • Each Midnight row's OnEnter / OnLeave / OnClick handlers are stripped once, so a secret value can never reach the shared GameTooltip via a bar tooltip.
  • Mode → metric map (matched by English mode label): Damage Done → DamageDone, DPS → DamageDone (per-second text), Damage Taken → DamageTaken, Healing Done → HealingDone, Absorbs → Absorbs, Deaths → Deaths. Modes with no C_DamageMeter source (Friendly Fire, Healing Taken, Overhealing, DOT/HOT Uptime, Activity, Threat) render blank.

• GUI_Main.lua — added local WOW_RETAIL / local WOW_RETAIL_MIDNIGHT upvalues (same expression as Recount.lua) and a single guarded early-return at the top of RefreshMainWindow: if WOW_RETAIL_MIDNIGHT and Recount.RefreshMainWindow_Midnight then return Recount:RefreshMainWindow_Midnight(datarefresh) end. On every other client this if is false and RefreshMainWindow runs byte-for-byte as before. Routing through the existing periodic refresh means the bars also keep updating between events. Also cached C_PetBattles as a file-local upvalue (was referenced unqualified at line 661).

• .luarc.json — added AbbreviateNumbers, LOCALIZED_CLASS_NAMES_MALE, canaccessvalue, and a batch of pre-existing WoW globals that were triggering undefined-global warnings across the project (BackdropTemplateMixin, C_PetBattles, UnitGUID, UnitHealth, UnitHealthMax, UnitIsPlayer, UnitIsConnected, UnitIsVisible, UnitAffectingCombat, UnitCanAttack, UnitDetailedThreatSituation, UnitIsFeignDeath, GetScreenHeight, UIFrameFade, GetMouseButtonClicked, GetFramerate, GetLocale, GetNetStats, GetNumDeclensionSets, GetChannelList, GetZonePVPInfo, BNGetNumFriends, BNGetSelectedFriend, Ambiguate, DeclineName, debugstack, strsplit, ChatThrottleLib).

• .pkgmeta — fixed the ignore entries so docs/ is actually excluded from the CurseForge zip. The packager appends /* to directory entries, so a trailing slash produced docs//* which matched nothing and silently shipped the folder. Removed the trailing slash from docs, .git, .github, .vscode; added .claude; documented the rule in a comment.

Multi-version safety

• Everything keys off WOW_RETAIL_MIDNIGHT (WOW_PROJECT_MAINLINE and build >= 120000). Classic Era, TBC, Wrath, Cata, MoP Classic, and pre-12.0 retail (11.x) are untouched — they never load the Midnight code path and RefreshMainWindow is unchanged for them.

Known limitations on Midnight

• During an instance encounter / Mythic+ / PvP match the values are passed through and display live; views Blizzard exposes no data for stay blank (listed above). No Total bar, no percentages, and bars are non-interactive (no click-to-detail / graph) on Midnight, because those paths need to read the underlying numbers or open a tooltip.

Files Changed

• Tracker_Midnight.lua — rewritten as the secret-safe display-passthrough renderer; old snapshot/diff AddAmount path removed
• GUI_Main.lua — WOW_RETAIL / WOW_RETAIL_MIDNIGHT upvalues, guarded RefreshMainWindow hook, C_PetBattles local
• .luarc.json — added the secret-value globals and a batch of pre-existing WoW globals to diagnostics.globals
• .pkgmeta — fixed docs/ (and other directory) ignore globs so docs are excluded from the package

[v1.18.2] (2026-05-16) - Bug fixes

Bug Fixes

• Prayer of Mending now credits the priest, not the unit being healed. Reported by a user: PoM heals were being attributed to the target rather than the healer. Root cause: Blizzard's combat log reports the buffed unit as the source of the SPELL_HEAL proc (spell ID 33110) — srcGUID and dstGUID are both the target — so Recount was attributing every PoM tick to whoever happened to be wearing the buff. Fix landed in Tracker.lua:
  • New module-local tables PoMAuraSpellId = {[41635] = true}, PoMHealSpellId = {[33110] = true}, and PoMOrigCaster = {}. The aura ID is the bouncing PoM buff; the heal ID is the heal-on-damage proc; the table is keyed by the GUID currently wearing the buff and stores {srcGUID, srcName, srcFlags} of the original priest.
  • Recount:SpellAuraApplied and Recount:SpellAuraRefresh write PoMOrigCaster[dstGUID] whenever a PoM aura lands or refreshes. The buff naturally rewrites the table entry as it bounces between targets (fresh SPELL_AURA_APPLIED events from the priest each hop).
  • Recount:SpellAuraRemoved clears PoMOrigCaster[dstGUID] when the buff falls off.
  • Recount:SpellHeal checks PoMHealSpellId[spellId] at the top of the function; if matched and PoMOrigCaster[srcGUID] exists, it swaps srcGUID, srcName, srcFlags to the recorded priest before dispatching to the existing AddHealData path. All downstream attribution (Healing Done, HealedWho, WhoHealed, overhealing, time-heal, etc.) now records the priest correctly with zero changes to the heal-tracking code itself.
  • PoM does not exist on Classic Era (Vanilla 1.x), so the lookups are a no-op there — no version guard required. Applies on TBC Classic, Wrath Classic, Cata Classic, MoP Classic, and Retail (where PoM uses the same spell IDs 41635 / 33110).

[v1.18.1] (2026-05-05) - Retail Midnight (12.0) support via C_DamageMeter

Bug Fixes

• Window color picker fixed on Classic Era and other Classic flavors. Reported by a user: "any chance of fixing the feature to change window colours?". Two bugs in one — both fixed:

  1. The picker code in colors.lua branched on WOW_RETAIL to choose between the new (Dragonflight-era) ColorPickerFrame.Content.ColorPicker API and the old OpacitySliderFrame / ColorPickerFrame.func API. But the new picker has since rolled out to every Classic flavor too — Vanilla Classic 11507/11508, Cata Classic, Wrath Classic, MoP Classic — leaving Classic clients in a half-migrated state where swatchFunc was set on Vanilla but the alpha read still went through the long-removed OpacitySliderFrame (which is nil). Replaced every WOW_RETAIL/WOW_VANILLA_CLASSIC/WOW_PANDA_CLASSIC flag check inside the picker with a runtime test for ColorPickerFrame.Content and ColorPickerFrame.Content.ColorPicker.

  2. Per-field assignment of swatchFunc / opacityFunc / cancelFunc on ColorPickerFrame followed by :Show() is silently dropped on the new picker — the OkayButton's OnClick reads from the info table the picker captured during setup, not from the frame's fields. Confirmed in-game: pressing OK on the new picker threw attempt to call field 'swatchFunc' (a nil value) from Blizzard's ColorPickerFrame.xml:79_OnClick. Refactored Colors:EditColor to use ColorPickerFrame:SetupColorPickerAndShow({r=, g=, b=, opacity=, hasOpacity=, swatchFunc=, opacityFunc=, cancelFunc=}) on the new picker (the documented Dragonflight+ pattern); the old per-field-then-Show path is preserved verbatim for any pre-Dragonflight client that still uses it. Position-relative-to-Attach moved to after Setup since the new picker re-anchors during its own setup.

  3. The first new-picker detector tested for ColorPickerFrame.Content.ColorPicker, but that field returned nil on at least Classic Era 11507/11508 even though the new-style OkayButton (which reads swatchFunc) was active — SetupColorPickerAndShow was never called and the second OK click reproduced the same error. Switched the detector to test for ColorPickerFrame.SetupColorPickerAndShow directly (the canonical 10.2.5+ method); if that method exists, the info-table contract is in force.

  4. Alpha reads in Color_Change and Opacity_Change previously reached into ColorPickerFrame.Content.ColorPicker:GetColorAlpha(). Switched to the documented top-level ColorPickerFrame:GetColorAlpha() so the same code works regardless of inner widget naming differences across Classic flavors.

  5. Color_Cancel no longer manually rewinds the picker's internal opacity slider — Colors:SetColor(Cur_Branch, Cur_Name, PreviousColor) already re-paints the registered visual elements via UpdateColor, and the picker frame is closing anyway. Removes the last ColorPickerFrame.Content.ColorPicker reference from the file.

  Also removed three now-unused project-flag locals from colors.lua (WOW_RETAIL, WOW_VANILLA_CLASSIC, WOW_PANDA_CLASSIC) and the dangling MoP-only swatchFunc = func post-Show fallback.

• No more ADDON_ACTION_FORBIDDEN Lua error at login on retail Midnight (12.0). A user reported 25x [ADDON_ACTION_FORBIDDEN] AddOn 'Recount' tried to call the protected function 'Frame:RegisterEvent()' traced to Recount.lua:1868 (the Recount.events:RegisterEvent("COMBAT_LOG_EVENT_UNFILTERED") call inside OnEnable). Root cause: in patch 12.0, Blizzard removed addon access to COMBAT_LOG_EVENT_UNFILTERED unconditionally. The call now fires ADDON_ACTION_FORBIDDEN in every context, including at login outside any combat / encounter / instance. Two prior research dead-ends ruled out before landing the real fix: (1) C_RestrictedActions.IsAddOnRestrictionActive(0..3) is not a useful gate — it reports "no restrictions" at login while the underlying call is still blocked. (2) pcall does not help — ADDON_ACTION_FORBIDDEN is dispatched asynchronously by Blizzard, not raised at the call site.

Retail Midnight scaffolding (data path NOT yet functional — tracked for v1.18.2)

• New file Tracker_Midnight.lua (loaded only on Mainline TOC, never on Classic flavors) is the intended CLEU-free data path. Subscribes to DAMAGE_METER_CURRENT_SESSION_UPDATED, DAMAGE_METER_COMBAT_SESSION_UPDATED, and DAMAGE_METER_RESET. On each session update, polls C_DamageMeter.GetCombatSessionSourceFromType(Current, metric, GUID) for every group member and would compute per-(player, datatype) deltas against a snapshot table, feeding them through the existing Recount:AddAmount(combatant, datatype, amount) so the rest of Recount (modes, fight history, bars, reports, etc.) works unchanged.
• In-game test outcome on retail Midnight (12.0.5.67314): source.totalAmount reads always come back as secret number values — not just during combat as the wiki's SecretWhenInCombat annotation suggested, but also after combat. Tainted addon code can hold a secret value but cannot do arithmetic on it ("attempt to perform arithmetic on local 'total' (a secret number value, while execution tainted by 'Recount')"). The first iteration of Tracker_Midnight.lua threw this Lua error every combat. The current iteration uses issecretvalue(total) to skip the metric and does NOT update the snapshot when the value is secret — eliminates the Lua error, but post-combat reads also come back secret in practice, so no deltas are ever applied and the Recount bars stay empty on retail Midnight.
• Net v1.18.1 effect on retail Midnight: addon loads cleanly, no Lua errors, settings / minimap / Titan Panel all work, but combat data does not populate. v1.18.2 will pursue the workaround — likely involving Skada's SecretValueHelper.lua pattern or finding a non-tainted code path the wiki documents elsewhere.
• Self-help warning kept in place: if C_DamageMeter.IsDamageMeterAvailable() returns false (in-game damage meter setting is OFF), Recount emits a yellow chat message at login telling the user to enable it under Options > Gameplay Enhancements > Damage Meter. (Won't matter until the secret-value problem is solved, but the warning is harmless and the setting is needed regardless.)

Implementation

• Recount.lua top-of-file — added local WOW_RETAIL = WOW_PROJECT_ID == WOW_PROJECT_MAINLINE (matching the convention in colors.lua) and local WOW_RETAIL_MIDNIGHT = WOW_RETAIL and ((select(4, GetBuildInfo()) or 0) >= 120000). GetBuildInfo and select cached as locals per project convention.

• Recount.lua new dispatcher Recount:RegisterCombatLogEvent — on Midnight, calls Recount:RegisterCombatLogEvent_Midnight() (defined in Tracker_Midnight.lua). On every other client, calls Recount.events:RegisterEvent("COMBAT_LOG_EVENT_UNFILTERED") exactly as before.

• Recount.lua OnEnable — line 1868 changed from Recount.events:RegisterEvent("COMBAT_LOG_EVENT_UNFILTERED") to Recount:RegisterCombatLogEvent(). The INSTANCE_ENCOUNTER_ENGAGE_UNIT registration on the next line is unchanged — that event is not subject to the CLEU restriction.

• Tracker_Midnight.lua (new file) — implements Recount:RegisterCombatLogEvent_Midnight(), the Recount:DAMAGE_METER_* event handlers, and the snapshot/diff polling. Uses synthesized COMBATLOG_OBJECT_* flags (Self/Party/Raid + Player + Friendly + ControlPlayer) when calling Recount:AddCombatant since real CLEU flags aren't available on this code path. Every C_DamageMeter call is wrapped in pcall, including the field access on the returned struct (the DamageMeterCombatSessionSource is marked SecretWhenInCombat).

• Secret-value handling — confirmed in-game on retail Midnight that source.totalAmount reads return secret number values during combat. Reading them does not throw, but arithmetic on them does ("attempt to perform arithmetic on local 'total' (a secret number value, while execution tainted by 'Recount')"). Worked around by checking issecretvalue(total) before any arithmetic and skipping the update when true. The snapshot is left alone in the secret case so the next non-secret read produces the correct delta against the last public total. Belt-and-suspenders pcall wraps the subtraction itself in case issecretvalue isn't reliable on a given build. Practical effect: bars don't tick up live during combat, but populate fully once the next post-combat session update fires.

• Recount_Mainline.toc — added Tracker_Midnight.lua immediately after Tracker.lua. The other four TOCs (Recount.toc for Classic Era, Recount_BCC.toc, Recount_Wrath.toc, Recount_Cata.toc) do NOT reference the new file, so Classic-flavor clients literally never load it. Zero risk of accidental cross-version interference.

• .luarc.json — added GetBuildInfo and C_DamageMeter to diagnostics.globals.

Behavior

• Pre-12.0 retail and all Classic flavors — no behavior change. CLEU registers immediately at OnEnable exactly as before. Tracker_Midnight.lua is not loaded.
• Midnight retail (12.0+) — login Lua error fixed; addon loads cleanly. Tracker_Midnight.lua registers the DAMAGE_METER_* events but in practice every poll returns secret values that we cannot operate on, so the Recount bars stay empty until v1.18.2 lands an actual workaround for secret values.
• No more login Lua error in any scenario, including on retail Midnight.

What's deferred to v1.18.2

• Get retail Midnight bars actually populating. The Tracker_Midnight.lua scaffolding is in place; the missing piece is a way to either (a) extract numeric values from secret-value returns, (b) call C_DamageMeter from a non-tainted execution path, or (c) display secret values directly via FontString widgets without the addon ever doing arithmetic. Likely path: study Skada's SecretValueHelper.lua (visible in zarnivoop/skada on GitHub) which is the only public reference for a working Midnight damage meter.
• Tracker modules on Midnight retail. Dispels, Interrupts, CC Breakers, Power Gains, Resurrections write to per-spell table-data structures (AddTableDataStats, AddTableDataSum) that don't have a 1:1 mapping to C_DamageMeter's combatSpells array. Will need module-specific shims once the basic data path works.
• Fight rotation alignment. Recount's fight boundaries come from PLAYER_REGEN_* (still works on Midnight), but Blizzard's session reset cadence may not match exactly. If the snapshots and Recount's CurrentFightData get out of sync at fight boundaries, follow-up tuning is required.

Files Changed

• Recount.lua — added Midnight version guard and the RegisterCombatLogEvent dispatcher; replaced the unconditional CLEU RegisterEvent call in OnEnable
• Tracker_Midnight.lua — new file, Midnight-only data collection via C_DamageMeter
• Recount_Mainline.toc — added Tracker_Midnight.lua to the load order
• colors.lua — replaced WOW_RETAIL / WOW_VANILLA_CLASSIC / WOW_PANDA_CLASSIC flag checks inside the color picker with a runtime test for the new ColorPickerFrame.Content.ColorPicker API; removed the three now-unused project-flag locals and the post-Show swatchFunc = func fallback
• .luarc.json — added GetBuildInfo, C_DamageMeter, and issecretvalue to diagnostics.globals

Older releases (v1.17.5 and earlier) are in CHANGELOG_ARCHIVE.md.