Changelog

All notable changes to ArenaCoachTBC are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

[2.8.40] - 2026-06-07

Fixed

• Tremor/Totem advice is now roster-gated. The engine suppresses CALL_TREMOR_FEAR, CALL_TREMOR_DOWN, and profile-driven CALL_SAVE_TREMOR_HOJ unless the friendly team has a living shaman or explicit Tremor capability.
• Opponent-profile notes no longer claim suppressed Tremor advice. Profile contribution logging only records a tendency when its matching callout was actually inserted.

Tests

• Added StrategyEngine regressions for no-shaman arena states and no-shaman opponent-profile states so impossible Tremor advice cannot leak back into the HUD.
• Local validation: 704/704 tests passing, locale parity green at 168 keys per locale, package-shape gate green, standalone StrategyEngine smoke green, and clean luacov total coverage is 99.11%.

[2.8.39] - 2026-06-07

Changed

• Shaman Bloodlust is now a personal burst action. When the burst gate opens and the player is the shaman, the live board and DBM-style alert show YOU: Bloodlust now instead of leaving the moment as a generic BURST NOW cue.
• Bloodlust is treated as a self/party action. Shaman lust advice no longer inherits the kill target; the kill plan stays in the strategy/context line while the shaman action stays clean and targetless.
• Shaman burst priority is urgent. The player-action stack now marks the Bloodlust row as urgent during a live burst window.

Tests

• Added StrategyEngine coverage for player-shaman burst windows producing an urgent targetless ACTION_SHAMAN_BLOODLUST.
• Added UI coverage for YOU: Bloodlust now taking the center alert while the generic burst cue remains supporting context.
• Local validation: 702/702 tests passing, locale parity green at 168 keys per locale, package-shape gate green, standalone StrategyEngine smoke green, and clean luacov total coverage is 99.10%.

[2.8.38] - 2026-06-07

Changed

• Live alerts now explicitly address the player. When the engine has a unit="player" action, the board and DBM-style alert render it as YOU: <action> so the center text is unmistakably about what the player should do next.
• The lower action strip is now a personal next-move surface. The visible header changed from generic assignments to "Your next move"; the player's row is simpler and louder, while teammate rows are dimmer supporting context.
• Strategy remains in the top strip. The top line continues to carry bracket, aggression, and the current plan so tactical context is available without competing with the player's action.

Tests

• Updated UI regression coverage for explicit YOU: self-action alerts and the personal next-move strip.
• Local validation: 700/700 tests passing, locale parity green at 168 keys per locale, package-shape gate green, standalone StrategyEngine smoke green, and clean luacov total coverage is 99.08%.

[2.8.37] - 2026-06-07

Changed

• HUD hierarchy now starts with the player's own action. When per-player assignments are available, the main board and DBM-style alert foreground what YOU should do now, while the global target/callout moves into supporting strategy context.
• The top strip is now strategy-first. The decorative Obsidian signature text was replaced with a live strategy summary showing bracket, aggression, and current plan.
• The side rail now keeps strategy/profile context persistent. It shows the current plan, aggression setting, comp/profile hints, and then only the highest-value tactical cues.
• Bottom assignments are now DBM-style action bars. Player and teammate actions stack top-to-bottom with the player's row first and a slow ten-second bar decay, making fight info readable long enough to capture.
• Stale fade is slower and less jumpy. Recommendations now wait longer before fading and fade over a longer window instead of disappearing almost immediately.

Tests

• Added UI coverage for player-action-first alerts, persistent strategy/profile rail content, and slow action-bar decay.
• Local validation: 700/700 tests passing, locale parity green at 168 keys per locale, package-shape gate green, standalone StrategyEngine smoke green, and clean luacov total coverage is 99.08%.

[2.8.36] - 2026-06-07

Fixed

• CI coverage gate now matches a clean GitHub runner. Added focused coverage for the new DBM-alert fallback paths, invalid display-mode fallback, detached module show path, version fallback, and burst-promoted detail line. This fixes the primary GitHub tests workflow failure from v2.8.35 without changing the live alert behavior.

Tests

• Clean local validation now mirrors CI: removed old luacov files before running the full suite.
• Local validation: 698/698 tests passing, locale parity green at 163 keys per locale, package-shape gate green, standalone StrategyEngine smoke green, and clean luacov total coverage is 99.03%.

[2.8.35] - 2026-06-07

Changed

• DBM-style center text now names the actual mechanic or action. The live alert and optional board no longer promote abstract mode words like KILL, DEFEND, 攻, or 守 as the big center line when a concrete cue exists. The primary line now reads like Purge Holyman, Tremor down - shaman refresh, BURST NOW, Kill Holyman, or Swap Frostbiter; mode remains only as color/context.
• Generic arcade buzz words were demoted. The small signal strip now stays as a quiet live marker instead of repeating vague state words such as DANGER or ATTACK.

Tests

• Added UI regression coverage that the default alert and board foreground concrete DBM-style action text instead of mode labels.
• Local validation: 696/696 tests passing, locale parity green at 163 keys per locale, package-shape gate green, and luacov total coverage is 99.55%.

[2.8.34] - 2026-06-07

Fixed

• Arena prep no longer shows live DEFEND from stale pressure. Before gates open, arena recommendations now stay in opener planning unless a static matchup pre-plan explicitly calls for a defensive start. Stale healerUnderPressure, old friendly damage samples, low HP, CC, or burst observations from a prior match can no longer paint DEFEND - healer trained before combat begins.
• Fresh arena entry clears volatile pressure state. PLAYER_ENTERING_WORLD now resets old healer-train observations and friendly damage timestamps alongside enemy identities and target history.

Tests

• Added StrategyEngine coverage for pre-gates arena state with stale healer pressure staying OPEN.
• Added Core reset coverage for clearing stale healer pressure and old damage samples on arena/world entry.
• Local validation: 694/694 tests passing, locale parity green at 157 keys per locale, package-shape gate green, and luacov total coverage is 99.50%.

[2.8.33] - 2026-06-07

Added

• Tremor-down shaman alert. When live aura scanning can see that Tremor Totem is missing while a fear threat is present, the compact HUD promotes Tremor down - shaman refresh to the top callout, plays the existing arena-gated warning cue, and gives the shaman an urgent Refresh Tremor Totem player action.

Fixed

• Tremor advice now clears from real state instead of lingering as generic comp text. Once the Tremor aura is detected again, the refresh callout and shaman assignment drop out on the next evaluation, letting the normal purge/grounding/kill advice return.

Tests

• Added StrategyEngine coverage for Tremor-down alert promotion and automatic clearing when Tremor is active.
• Added Core aura-scan coverage for Tremor Totem through both UnitAura and UnitBuff paths.
• Local validation: 693/693 tests passing, locale parity green at 157 keys per locale, package-shape gate green, and luacov total coverage is 99.46%.

[2.8.32] - 2026-06-07

Changed

• Default live HUD is now a DBM-style alert instead of the full fixed board. ArenaCoachTBC now shows a compact movable center warning such as !! KILL !! Holyman, !! SWAP !! Mage, or !! DEFEND !!, with the top actionable callout on a subline and the existing arena-gated SoundKit cues for mode/callout changes.
• The Obsidian board is now opt-in for review/tuning. Use /acc hud board to show the full divided board or /acc hud both to compare the alert and board together. /acc hud alert returns to the live default and hides the board immediately.

Fixed

• Stale fade works in alert-only mode. The DBM-style alert participates in the same stale-recommendation fade timer as the board, then clears nameplate and edge cues when the situation is out of sync.

Tests

• Added UI and slash-command coverage for alert creation, default alert rendering, board suppression, board mode, stale alert fade, and /acc hud alert|board|both.
• Local validation: 690/690 tests passing, locale parity green at 155 keys per locale, package-shape gate green, and luacov total coverage is 99.37%.

[2.8.31] - 2026-06-07

Fixed

• CurseForge release uploads now fail loudly instead of disappearing. Stable-tag releases no longer mark the BigWigs/CurseForge upload step continue-on-error, so a rejected or failed upload blocks the release workflow instead of leaving GitHub green while CurseForge has no usable file.
• BCC game-version upload is forced for CurseForge. The BigWigs packager now runs with -g bcc, using ## Interface-BCC: 20505 so the file is classified for TBC/BCC Anniversary (2.5.5) rather than relying on automatic detection.

Tests

• Added tools/check_package_shape.lua and wired it into CI. The gate verifies TOC file existence, shipped-file count, version consistency, CurseForge project metadata, .pkgmeta inline-comment safety, the GitHub zip-shape validation step, and the non-optional BigWigs upload step.
• Release workflow now validates the GitHub zip before publishing: it must contain ArenaCoachTBC/ArenaCoachTBC.toc, include more than 20 non-empty addon files, and exclude Tests/ plus dotfiles.
• Performance budget specs now keep the original raw budgets while applying luacov-specific slack, preventing coverage hook overhead from causing unrelated release flakes.
• Local validation: 687/687 tests passing, locale parity green at 154 keys per locale, package-shape gate green, local BigWigs dry-run produced a full ArenaCoachTBC-v2.8.31-bcc.zip, and luacov total coverage is 99.06%.

[2.8.30] - 2026-05-27

Fixed

• Targetless RESET no longer pop/fades the HUD. Live arena/BG/world contexts now keep the frame hidden when the engine has no actionable target, instead of briefly showing a low-value RESET frame and letting stale fade remove it.

[2.8.29] - 2026-05-27

Fixed

• PvE mobs no longer wake the world-PvP HUD. World/BG CLEU fallback now requires a real player GUID before refreshing the hostile-contact timer or creating a non-arena enemy stub, so hitting or being hit by ordinary creatures while PvP-flagged leaves the HUD hidden.

[2.8.28] - 2026-05-27

Changed

• Bottom assignments now prioritize your own action. The player assignment is promoted to the first bottom card when needed, marked with a bright YOU tag, and rendered on a stronger plate so the answer to "what do I do right now?" is visible before team context.
• Assignment cards are denser and clearer. Bottom cards now use three scan-friendly lines instead of cramped four-line blocks, keeping action text readable inside the compact HUD.

[2.8.27] - 2026-05-27

Fixed

• HUD readability on busy backgrounds. Strengthened the main shell, child-panel reading plates, drag strip, and assignment cards so text no longer fights dense world geometry.
• Side rail headers no longer wrap. Shortened focus and cue subtitles to fit compact rails without splitting words across lines.
• DEFEND/RESET no longer show an inactive health strip. The center health bar and label now hide whenever the current mode is not target-attached, preventing overlap with defensive advice.

[2.8.26] - 2026-05-27

Changed

• Obsidian Signal visual language. The HUD now uses warmed obsidian surfaces, burnished brass rules, cyan intelligence accents, bone-white data text, and restrained crimson signal colour.
• Surveyor reticles added. Compact brass corner marks and a persistent OBSIDIAN / SIGNAL metadata strip make the board read more like a deliberate tactical instrument.
• Health and probability colours tightened. Kill probability and target health now use the Obsidian Signal palette instead of generic green/yellow/red UI colours.

[2.8.25] - 2026-05-27

Changed

• Bottom assignment strip now divides into 1/2/3/5 slots. The player-info row uses fixed small cards based on current player actions, alive friendlies, or arena bracket instead of a loose paragraph block.
• 2v2, 3v3, and 5v5 jobs keep stable positions. A five-player recommendation now fills five compact cards in normal mode, while 2-player advice hides unused cells.
• Assignment header stays visible. The bottom row keeps a tactical slot label so the divided area is easy to read and drag around during HUD layout checks.

[2.8.24] - 2026-05-27

Changed

• Tactical-console visual pass. Added a top signal strip, ruler ticks, health-pool bar, console-style section headers, and row/card text treatments inspired by the reference cockpit mockup.
• Center panel now reads as an instrument. The action call keeps the target/stats text, while the health bar shows current target HP at a glance without adding flashing effects.
• Focus, cue, and assignment rows are more scannable. Rows now use compact tags, muted subtitles, and priority-coloured target text instead of plain paragraph labels.

[2.8.23] - 2026-05-27

Changed

• Prototype-A now follows the original cockpit sketch more closely. The HUD uses a persistent left status stack, center action panel, center player-info/assignment panel, and right cue rail instead of equal-width top boxes plus a full-width assignment row.
• Side rails now span the board height. Focus and cue sections stay visually stable while the center column handles the active call and per-player jobs.
• Mode accent bar added. The center action panel gets a slim mode-coloured accent line so KILL/SWAP/DEFEND state is readable without screen flashing.

[2.8.22] - 2026-05-27

Fixed

• Prototype-A sections no longer collide. Added gutters, a reserved assignment row, and stronger minimum-height bounds so the focus, action, cue, and assignment regions stay visually separated.
• Center detail is capped to the action section. Verbose comp, chain, and callout lines are now limited by available center height so they cannot spill into the assignment row.
• Verbose detail scales with board size. Compact boards clip assignments and cues for readability, while taller/wider resized boards can show the full five-player assignment list for review.

[2.8.21] - 2026-05-27

Fixed

• Master off now wins over every HUD path. /acc off keeps newly-created frames hidden, prevents UI:Show / /acc toggle from reopening the HUD, and blocks forced test/simulator recommendations from painting the GUI while disabled.
• Manual simulations stop on disable. /acc off now cancels active simulator callbacks so a running /acc test or /acc simulate cannot keep feeding stale visual beats after the user turns the addon off.
• Options-panel disable matches slash disable. Turning off the Enabled checkbox now hides the HUD, edge cue, and nameplate paint immediately instead of only flipping the SavedVariables flag.

[2.8.20] - 2026-05-27

Changed

• Prototype-A board is resizable. /acc unlock now exposes a lower-right grip on the integrated HUD board, with width/height saved alongside position and clamped to arena-safe bounds.
• HUD zones reflow with the board. The left focus, center action, right cue, and lower assignment panels resize together so enlarging the HUD gives text more room without breaking the single-box layout.

[2.8.19] - 2026-05-27

Changed

• HUD board is lighter. Reduced the main board, module panels, and drag-strip opacity so the playfield remains visible behind the prototype-A layout.
• Text carries the contrast now. Added stronger shadowing to the title, drag grip, center action, stats, callout, focus, cue, and assignment text so the lighter board remains readable without a dark screen-covering panel.

[2.8.18] - 2026-05-27

Fixed

• Timed /acc test now repaints the HUD on every scheduled beat. The simulator now drives a simulator-owned engine tick that calls StrategyEngine:Evaluate, force-shows the HUD for out-of-arena tests, publishes the WeakAura payload, and reports any simulation/HUD error to chat. This prevents the live C_Timer.After replay from printing scenario lines while the board remains stuck on waiting placeholders.

[2.8.17] - 2026-05-27

Fixed

• Prototype-A now has an obvious draggable shell. The integrated HUD uses a stronger translucent background, darker top drag strip, grip marker, and visible internal dividers so the left focus, center action, right cue, and lower assignments zones start inside one clear box instead of reading as loose floating text.
• Detached prototype modules reset off for migrated layouts. Older SavedVariables that accidentally enabled satellite modules are migrated back to the integrated board default, while the legacy detached frames remain dormant unless deliberately re-enabled.

[2.8.16] - 2026-05-27

Fixed

• Manual tests cannot silently run while disabled. /acc test and /acc simulate <scenario> now turn the master switch back on before replaying. Previously, if ArenaCoachTBCDB.enabled=false, chat scenario events continued but Core:Evaluate() returned early, leaving the HUD stuck on the initial waiting text.

[2.8.15] - 2026-05-27

Fixed

• Prototype-A is now visibly one board. The HUD no longer depends on detached satellite frames to communicate the design. The main frame itself contains a texture-backed left focus panel, center action panel, right cue panel, and lower assignment panel, so moving the HUD cannot leave the design pieces behind.
• Backdrop-independent panels. The board and panels now use explicit texture backgrounds/borders instead of relying only on SetBackdrop, making the layout visible on clients where the old backdrop path renders as plain floating text.

[2.8.14] - 2026-05-27

Fixed

• Prototype-A layout is visible while waiting. The HUD no longer collapses into only floating center text before the fight starts or during /acc test hud. The left focus strip, center action toast, right cue rail, and lower assignments module now show structural waiting placeholders until real target/cue/assignment data arrives.
• 640-wide arena fit. The center toast is now 300x118, side modules are 150px wide, and untouched v2.8.13 side-module SavedVariables migrate inward so the layout reads as left / center / right in compact screenshots instead of drifting off the edges.

Tests

• Added regression coverage for force-shown HUD demos, pre-gate scaffolding, initial waiting scaffolding, and default-position migration.

[2.8.13] - 2026-05-27

Changed

• Prototype-A HUD module layout. The built-in HUD now matches the agreed layout: left focus strip, center action toast, right cue/icon rail, and lower assignment module. /acc unlock makes all four modules independently draggable; /acc lock locks them.
• Focus and cue side modules. The left strip summarizes the primary target, swap candidate, and lowest friendly pressure when known. The right rail renders callout icons/text for burst, purge, HoJ, peel, dispel, and other top cues.

[2.8.12] - 2026-05-27

Changed

• Movable assignment module. Player assignments now render in their own compact frame instead of being welded to the main action toast. /acc unlock lets the main call and assignments be dragged independently; /acc lock locks both. The new position persists in db.assignmentFrame.
• Smaller main action toast. The live action toast is now 320x118 with slightly smaller arcade/action/stat text and a quieter backdrop, leaving more room for arena frames, nameplates, cast bars, action bars, DBM bars, WeakAuras, chat, and damage meters.

[2.8.11] - 2026-05-27

Changed

• Compact prototype-A HUD. The live frame is now a 340x168 action toast instead of a large center board: smaller arcade cue, smaller main action text, tighter stats/reason rows, quieter backdrop, and up to three default assignment lines so a 3v3 team still gets per-player advice without covering other addons.
• Faster stale fade. Out-of-sync recommendations now begin fading after 2.5 seconds and finish after 1.5 seconds, clearing stale nameplate/edge cues with the frame.

Fixed

• /acc test hud subcommand routing. Slash-command input now forwards the hud, bg, world, and print arguments into RunTestMode, so the visual HUD tour can be launched from chat instead of always falling back to the default realistic replay.

[2.8.10] - 2026-05-27

Changed

• Readable /acc test pacing. The realistic arena replay now spans about a minute instead of compressing an entire match into a rapid 11-second sequence. Bookkeeping events can update scenario state without forcing redundant HUD repaints, so the replay reads like an arena opener instead of a flashing demo.

[2.8.9] - 2026-05-27

Fixed

• Combat-log payload compatibility. COMBAT_LOG_EVENT_UNFILTERED parsing now accepts both modern CombatLogGetCurrentEventInfo() and older vararg-style combat-log payloads. SWING_* events are parsed without pretending the damage amount is a spell ID, while SPELL_* events still expose spell id/name for trinkets, cooldowns, DR, and spec hints.

[2.8.8] - 2026-05-27

Added

• Realistic /acc test arena replay. The default /acc test path now runs a real engine-driven 3v3 RMP scenario instead of only forcing canned HUD recommendations. It simulates pre-gates OPEN, combat start, enemy burst, healer CC, train pressure into DEFEND, Disc Priest discovery through Pain Suppression, trinket/defensive state, a kill/swap window, and match reset.
• Simulator PvP context and friendly-pressure events. Built-in simulations can now seed bracket/context, friendlies, friendly health/debuffs, enemy debuffs, damage-pressure rings, phase changes, observations, and cleanup after the replay.

Changed

• Visual-only HUD tour moved to /acc test hud. The old DBM-style forced HUD walk-through remains available for checking arcade warning text, target stats, assignments, nameplate paint, and optional edge cues.

[2.8.7] - 2026-05-27

Added

• HUD version marker. The main ArenaCoachTBC frame now shows the loaded addon version in the top-right corner so rapid local-copy and release verification is visible in-game.

[2.8.6] - 2026-05-26

Fixed

• Real-fight damage pressure detection. Classic combat logs can report melee/contact hits as SWING_DAMAGE_LANDED and shield/contact damage as DAMAGE_SHIELD; those now count as damage pressure for BG/world enemy stubs and healer-train detection.

Docs

• Added a real-fight capture checklist covering /acc trace, /acc record, /combatlog, SavedVariables, and replay comparison.

[2.8.5] - 2026-05-26

Fixed

• Rated-arena decision quality pass. Corrected 2v2 pre-gate double-DPS/hybrid opener handling, strengthened low-HP 2v2 kill-window priority, added data-driven active kill targets for matchups such as WLP drain, restored target HP display from healthPct, formatted target-aware HUD callouts without raw %s, and gated BURST_NOW on the full BurstDecision.

Tests

• Added rated-arena regression coverage for the known benchmark misses and raised the benchmark floor to 85%.

Notes

• Merged on top of v2.8.4, preserving player assignments, stale HUD fade, arcade cues, subtle edge visuals, and distribution-copy updates.
• Tests 660/660 passing. Locale parity green at 145 keys per locale. Local luacov total coverage: 99.32%. Rated-arena benchmark agreement: 21/21.

[2.8.4] - 2026-05-26

Changed

• CurseForge project description rewritten for approval clarity. The dashboard copy now explicitly explains the features offered, what players experience in arena/BG/world PvP/duels, safety limits, privacy behavior, slash commands, localization, and current validation numbers.

Notes

• Docs-only distribution polish. Package version bumped so GitHub release notes and downloadable addon metadata stay aligned.

[2.8.3] - 2026-05-26

Added

• Stale HUD fade-out. Fresh recommendations restore the HUD to full opacity, but if the fight state stops refreshing for several seconds the central text fades away and hides instead of leaving an out-of-sync call on screen.

Fixed

• Stale visual layers clear with the fade. When the HUD fades out, optional edge cues and nameplate highlights are also cleared so target-specific advice cannot linger after the situation has moved on.

Notes

• Tests 636 -> 639. Locale parity green at 144 keys per locale. Local luacov total coverage: 99.10%.

[2.8.2] - 2026-05-26

Changed

• Removed the big flashing screen-edge feel. The optional edge visual is no longer a 96px pulsing band around the screen. Even if an older SavedVariables file has alerts.edgeGlow = true, it now renders as a thin 18px, low-alpha, static edge cue.
• HUD + docs now treat the arcade plate/nameplate as the primary visual warning. /acc glow on|off remains for users who want a subtle peripheral cue, but the default and documented experience avoids big screen-border motion.

Notes

• Added regression coverage that locks the edge cue to thin, low-alpha, and non-pulsing.
• Tests 635 -> 636. Locale parity green at 144 keys per locale. Local luacov total coverage: 99.10%.

[2.8.1] - 2026-05-26

Added

• Arcade warning plate. The HUD now renders a large passive warning cue above the tactical line, using high-impact arcade words such as READY, ATTACK, SWITCH, DANGER, BURST, and PINCH so urgent arena/BG/world PvP states are easier to parse at a glance without returning to fullscreen flashing.

Fixed

• Burst and outnumbered warnings are visually louder but still non-intrusive. BURST_NOW and outnumbered disengage states now promote to the arcade cue line while continuing to avoid screen flashes, protected actions, or chat automation.

Notes

• Tests 633 -> 635. Locale parity green at 144 keys per locale. Local luacov total coverage: 99.04%.

[2.8.0] - 2026-05-26

Added

• DBM-style per-player assignments. StrategyEngine:Evaluate now publishes rec.playerActions, one compact action per living friendly, with unit/name/class/action/target fields. The built-in HUD renders the assignment block under the main recommendation so a 3v3/5v5 team can see who should MS, purge, HoJ, peel, dispel, or reset.
• WeakAura bridge support for assignments. Added GetPlayerActions(), GetPlayerAction(), and GetActionForUnit(unit) so custom WeakAuras can render each player's assignment or only the local player's action.

Fixed

• Demo and locale text no longer mention DEFEND flashing. The no-flash behavior remains intact from v2.7.5.

Notes

• Tests 627 -> 633. Locale parity green at 135 keys per locale. Local luacov total coverage: 99.03%.

[2.7.6] - 2026-05-26

Fixed

• CI parity follow-up for the real-arena test suite. Preserved nil holes in the mocked CombatLogGetCurrentEventInfo() varargs so LuaJIT sees the same CLEU payload shape as Lua 5.1 and the WoW client.
• Coverage-aware performance gate. The raw StrategyEngine:Evaluate timing budget remains 5 ms, while the luacov-instrumented CI run now uses a 15 ms ceiling so coverage hooks do not masquerade as engine latency.
• Slash-command coverage for live visual toggles. Added regression coverage for /acc highcontrast, /acc verbose, /acc on|off, /acc glow, and /acc nameplate.

Notes

• Tests 626 -> 627. Local luacov total coverage: 99.12%.

[2.7.5] - 2026-05-26

Fixed

• Removed automatic full-screen flashing from live recommendations and /acc test. UI:Apply no longer calls the red _Flash() overlay for URGENT/DEFEND, even if an older SavedVariables file still has alerts.screenFlash = true. The quieter cues remain: HUD colour, nameplate highlight, optional edge glow, and arena-gated sound.
• Healer-train damage now re-evaluates through the real CLEU path once the peel threshold is reached. Repeated damage on a healer in arena now publishes DEFEND immediately instead of waiting for the next spellcast/aura event.

Added

• Realistic arena lifecycle regressions. Added event-driven tests that go through PLAYER_ENTERING_WORLD, ARENA_OPPONENT_UPDATE, PLAYER_REGEN_DISABLED, live arenaN unit stubs, CLEU healer damage, UI:Apply, and WeakAuraBridge publication. These cover the gates-closed OPEN state, active-combat KILL, healer-train DEFEND, and the no-flash behavior in a shape much closer to an actual arena run.

Notes

• Tests 624 → 626 (+2 real-arena lifecycle regressions). Locale parity green at 112 keys per locale. Full syntax check green.

[2.7.4] - 2026-05-26

Fixed

• Strategy logic pass for real PvP contexts. BurstDecision is now the single source of truth for BURST_NOW: target immunity, configured MS, Windfury, melee uptime, kill probability, chain readiness, and incoming pressure all land in the auditable gate table. The old separate burst prerequisite path was removed so HUD callouts and bridge API cannot disagree. Chain readiness is advisory by default and can be made strict with strategy.requireChainForBurst = true.
• DEFEND now works for support-capable teams and solo world PvP. Low Paladin/Shaman support friendlies count as defensive anchors, healer CC checks use healer/support capability instead of Priest/Druid-only checks, and solo world PvP falls back to the lowest alive friendly so a dying non-healer player can still get DEFEND.
• Non-arena enemy discovery is less noisy and less brittle. BG/world nameplate scanning no longer stops at a missing nameplate1, and CLEU fallback stubs are only created when a hostile source damages the player or a known friendly.
• Target and comp edge cases are safer. Immune/unreachable-only targets now produce RESET instead of a bad KILL call, primaryTargetHp publishes correctly from healthPct, low-mana healer kill probability works without roleGuess, and TRIPLE_DPS no longer matches 2-player double-DPS states.
• Profile contribution traces no longer duplicate trinketsFear.

Changed

• Updated the strategy/architecture docs to describe arena, BG, world PvP, burst gates, non-arena discovery, and the narrower dynamic TRIPLE_DPS fallback.

Notes

• Tests 614 → 624 (+10 focused regressions). Locale parity green at 112 keys per locale. Full syntax check green.
• Coverage was not run locally because luacov is not installed in this WSL environment; the CI coverage gate remains unchanged.

[2.7.3] - 2026-05-26

Fixed

• v2.7.1 outnumbered override was too aggressive — suppressed DEFEND in salvageable 2v3 / 1v2 emergencies. Found by a Codex adversarial review run against v2.6.0..HEAD: in a 3v3 down to 2v3 with the healer at 20% HP and pressure detected, the engine returned KILL + the CALL_OUTNUMBERED_DISENGAGE callout instead of DEFEND. The 1.5x ratio caught every state where you'd lost one friendly, not just the unrecoverable 2v4 case the override was designed for. Defensive cooldowns CAN save a 2v3; suppressing DEFEND there is a regression.

  Fix has two parts:

  1. Restructured shouldDefend so real-emergency signals check FIRST. low_healer (any healer below the HP threshold), healer_cc (healer mid-CC), enemy_lust (Bloodlust active), and multi_burst now short-circuit before the outnumbered override runs. These signals indicate that defensives are exactly what saves the team, regardless of numbers.
  2. Narrowed the outnumbered threshold from nEnemy >= nFriendly * 1.5 to nEnemy >= 4 AND (nEnemy - nFriendly) >= 2. Catches the original 2v4 / 2v5 / 1v3+ cases; no longer fires on 2v3 / 1v2 where defensives still work. isOutnumbered (the callout helper) follows the same threshold so the override + callout stay in sync.

Notes

• Tests 612 → 614 (+2 Codex-suggested regressions: arena 2v3 with low healer must still DEFEND; 2v4 with low healer alive at 15% must still DEFEND via low_healer taking precedence over outnumbered). Locale parity green.
• The original v2.7.1 user case (arena 2v4 with healerUnderPressure but no specific defensive signal) still works correctly: low_healer doesn't fire (nobody's low), enemy_lust doesn't fire, so we fall through to the outnumbered branch which suppresses DEFEND and adds the CALL_OUTNUMBERED_DISENGAGE callout.

[2.7.2] - 2026-05-26

Two stacked lifecycle bugs both rooted in stale state.

Fixed

• Engine recommended KILL before arena gates opened. User report: "it suggest to kill even before the game." Root cause: onArenaOpponentUpdate set combatPhase = "ACTIVE" the moment any opposing player became visible to the client — which fires while you're still in the prep room before gates open. The legitimate PRE → ACTIVE transition is PLAYER_REGEN_DISABLED (combat starts); that already handles it. Removed the spurious assignment so the pre-gates window correctly stays in OPEN mode.
• Stale phantom enemies from the previous match leaked into the next. User report: "always 15% showing, sometime a player name doesn't even exist." The 15% was the engine's baseline kill-probability fallback when no fresh data exists; the phantom name was a dead enemy from a prior arena that state.enemies never cleared. Now onPlayerEnteringWorld explicitly resets state.enemies, state.enemyClassList, state.lastPrimaryGUID, and state.combatPhase = "PRE" on every zone transition. CLEU + UnitAura subscriptions rebuild the state within one evaluation tick once combat starts, so the reload-mid-fight case recovers cleanly.

Notes

• Tests 611 → 612 (+2 regressions: ARENA_OPPONENT_UPDATE must NOT flip combatPhase to ACTIVE; PEW must reset per-match state including phantom enemy entries). Locale parity green.
• Engine, bridge API, storage shape unchanged. Pure lifecycle correctness fix.

[2.7.1] - 2026-05-26

Fixed

• Engine recommended DEFEND in a 2v4 arena. User report: "we got 2v4 situation, you ask me to defend." Root cause: shouldDefend() returned true via the "healer being trained" branch whenever multiple damage events landed on the healer in a 5-second window — which fires by definition in a 2v4 because 4 enemies attacking 2 players means multi-source damage. But defensive cooldowns can't save a 2v4: you spend them in one global and everyone dies. The actionable advice in that state is "disengage or counter-burst the lowest-HP enemy", not "burn Pain Sup".

  Added an outnumbered override at the top of shouldDefend(state): when alive enemies ≥ alive friendlies × 1.5 in arena context, suppress DEFEND and let decideMode fall through to KILL. A new CALL_OUTNUMBERED_DISENGAGE callout is added at the top of the callout list (so it gets the prominent icon + text slot in the HUD), localised in both locales: "Outnumbered — disengage or burst lowest-HP" / "敌众我寡 - 脱离或集火残血". Wired to the Aspect of the Cheetah icon (spell 5118) so the visual cue reads as "run away".

Notes

• Arena-only. The override does NOT apply in BG / world context. In BG the engine's "alive enemies" comes from nameplate scans and includes everyone in range, not just active combatants — a 3v10 nameplate count isn't a real outnumbered state. Arena's arenaN unit IDs are exactly the opposing team, so the ratio is meaningful there.
• 609 → 611 tests (+2 regression: arena 2v4 suppresses DEFEND + adds the callout; BG 3v10 still allows DEFEND).
• Locale parity green: 112 keys per locale (was 111; added CALL_OUTNUMBERED_DISENGAGE).
• Engine, bridge API, storage shape all unchanged. UI HUD adds the icon automatically via the v2.7.0 callout-icon map.

[2.7.0] - 2026-05-25

Visual hierarchy pass driven by user feedback. "Whole-screen glow doesn't help, you should add which role does what in the HUD, can you add some ICONs?" Two changes:

Changed

• Edge glow flipped from default-on to default-off. The full-screen pulsing band was more distraction than information per real-use feedback. Still available via /acc glow on if you want it back. Nameplate highlight stays default-on — that one's anchored to the actual kill / swap target so it carries role information, not just mode colour.
• Callouts now show their spell icon inline. Pre-v2.7 every callout rendered as ▸ HoJ kill target — text only, leaving you to translate "HoJ" → "the paladin's Hammer of Justice" → "which icon is that on my bars" in your head mid-fight. Now: |TInterface/Icons/Spell_Holy_HammerOfJustice|t HoJ kill target — the spell's actual in-game icon renders inline as a 18px texture, so you see the action visually. Mapping covers all current callouts (HoJ, Tremor, Grounding Totem, Purge, Dispel Magic, Pain Suppression, BoP, Cyclone, Psychic Scream, Mana Burn, Ice Block warning, Counterspell, BG flag carrier, Divine Shield, Bloodlust for BURST NOW, …). When GetSpellTexture returns nil (very first call on an unknown spell ID), the row degrades gracefully to the previous ▸ <text> bullet.

Notes

• This makes verbose mode (/acc verbose on) much more useful too — each callout in the list now reads as a stacked action menu with icons + text, not a pipe-separated text blob.
• Existing db.alerts.edgeGlow setting respected: users who explicitly turned it ON (or never edited it pre-v2.7) keep their current state. Only fresh installs see edge glow off.
• 609 tests still passing. Locale parity 111/111. Bridge API unchanged.

[2.6.0] - 2026-05-25

True closure of the v1 roadmap, plus user-feedback polish. Picks up the last 3 deferred-but-doable items (per-callout cooldown already shipped in v2.5.0, public wiki, LuaJIT CI matrix) plus two new user reports from v2.5.0 testing (demo too fast to read, HUD still feels cluttered).

Added

• LuaJIT 2.1 CI matrix (M6). .github/workflows/test.yml gained a parallel luajit-tests job that runs the full suite under LuaJIT 2.1. Lua 5.1 stays the contractual primary (TBC client uses 5.1 exclusively). The LuaJIT job is continue-on-error: true so it reports without blocking PRs — its purpose is regression-catching, not gating. LuaJIT 2.0 not added; apt-get on Ubuntu LTS only ships 2.1.
• Per-comp strategy primer wiki (M5). New docs/strategies/ directory with README.md (structure + framework) and rmp.md (the starter primer — full game plan, kill conditions, per-archetype variations, callout list, common mistakes). Other primers (WMS / TSG / Jungle / RLS / DRAIN / BG cleave) are stubs ready for contributor PRs.
• Demo slowdown. /acc test beats now space at 3 s instead of 2 s (1.5x multiplier; total demo 14 s → 21 s) so each beat is readable before the next replaces it. New /acc test slow keyword bumps to 2.5x (35 s total) for screen-share / streaming demos. End-of-demo restore delay also scales.

Changed

• HUD visual hierarchy polish.
  • f.statsText font bumped from GameFontHighlight (~12pt) to 18pt OUTLINE — readable at a glance, not just under careful inspection.
  • Stats segments now colour-coded inline: HP white (neutral reference value), kill prob green / amber / red (≥60 / 30-59 / <30), ★ BURST READY in gold with a leading sigil so the burst signal pops as the most attention-grabbing element on the line.
  • Wider segment separator ( · instead of ) and a leading sigil so segments breathe.
  • Vertical spacing between mode label / stats / sub-text widened from -2px / -4px to -8px / -8px so the sections read as distinct rows.
  • f.subText:SetSpacing(3) so multi-line text (verbose mode chain steps) doesn't crowd together.

Roadmap

• Marked these items DONE on ROADMAP.md:
  • Per-callout cooldown (was already shipped in v2.5.0, just unchecked)
  • Public wiki (docs/strategies/ covers this — README + starter primer; contributors add more)
  • CI matrix LuaJIT 2.1 (added; 2.0 not available via apt)
• The remaining unchecked items are the genuinely external-only ones: additional locales (need native speakers), cloud telemetry (principle conflict), dyslexia font (licensing), app icon/screenshots (design assets), web visualiser (separate project). These won't ship without external resources.

Notes

• Tests 609 still passing. Locale parity 111/111. Bridge API + engine surface unchanged.
• The v1 ROADMAP is now genuinely closed. The v2.x line is feature-complete pending external-resource items.

[2.5.0] - 2026-05-25

Polish release — closes out the v2.x line. Picks up the last three actionable items from the v1 roadmap that hadn't shipped yet: per-callout cooldown (M4), high-contrast accessibility skin (M4), and a tightened performance budget assertion (M6). ROADMAP.md updated to mark the remaining items as either shipped, deferred with explicit reasons (external dependencies — alternate Lua runtimes, native-speaker contributors, design assets, hosting infra), or permanently out of scope (cloud telemetry conflicts with operating principle #5).

Added

• Per-callout cooldown (M4). UI:Apply now tracks last-shown-time per callout key and suppresses the same callout for 3 seconds. Stops the "same text every 0.5 s" pattern that could surface if engine state oscillates around a threshold (enemy HP bouncing across the 50% gate, for example). Applies in both Quiet HUD and verbose modes.
• High-contrast HUD skin (M4 accessibility). New /acc highcontrast on|off (alias /acc hc) flips between the default visually-coherent palette and a fully-saturated primary palette (pure red KILL, pure yellow OPEN, pure orange SWAP, saturated cyan-blue DEFEND, white RESET). Persists in db.frame.highContrast. Useful on small screens, under glare, or for users with reduced colour sensitivity. The mode label colour swap is immediately repainted via a synthetic Evaluate when the toggle is flipped.
• Full-cycle perf budget assertion (M6). Tests/Performance_spec.lua gains a new test that exercises the full hot path — SE:Evaluate → UI:Apply → WeakAuraBridge:Publish — and asserts <15 ms mean over 100 iterations. The v2.2.5 city-lag bug came from onNameplateChange running this exact path on every nameplate event in world_idle; this budget cap means any future regression of that pattern will fail CI immediately.

Changed

• ROADMAP.md — M6 section restructured to distinguish (a) what shipped (per-Evaluate budget, lookahead+patterns budget, AV-scale 40-enemy budget, 100-arena memory fuzz, full-cycle budget), (b) what's deferred due to external dependencies (LuaJIT CI matrix, web visualiser, interactive replay UI), and (c) what's permanently out of scope (cloud telemetry, single-developer pseudo-locales, alternate fonts). The v1 ROADMAP is now a retrospective in steady state.

Notes

• 608 → 609 tests (one new full-cycle perf assertion). Locale parity green (still 111 keys per locale; no new locale work).
• Bridge API, engine, and storage shape all unchanged. No SavedVariables migration needed; db.frame.highContrast and db.frame.verbose (v2.4.0) auto-merge on next login.
• This release closes out the v1 ROADMAP and (along with v2.0.0 closing ROADMAP-v2) leaves the project in a feature-complete steady state. Future work, if any, falls into patch releases (bug fixes, new comp catalog entries) or a v3.0+ engine evolution.

[2.4.0] - 2026-05-25

Quiet HUD. Information density on the recommendation frame had grown to 5-6 lines of text per evaluation — too dense to parse mid-fight. User screenshots showed the wall-of-text problem in zhCN clients (where some keys still rendered as raw identifiers). v2.4 cuts the default HUD to two lines + the mode badge + the target stats row, moves the rest behind a /acc verbose toggle, and patches the last untranslated callout.

Fixed

• BURST_NOW callout rendered as the raw key in non-English clients ("BURST_NOW | 无敌锤上焦点" in the user's zh screenshot). Added the key to both Locales/enUS.lua ("BURST NOW") and Locales/zhCN.lua ("立即爆发"). Both locales now at 111 keys, parity green.

Changed

• HUD subText cut to one callout line in default mode. Pre-v2.4 every evaluation rendered: [reasonKey] | [callout1 | callout2 | callout3] | [comp badge] | [chain title (62%)] | [step 1] | [step 2] | [step 3] — six to seven lines mid-fight. v2.4 default shows just ▸ [top callout] (plus the localised reasonKey for DEFEND/RESET). The big mode label + target name + target stats row (HP%/kill prob/BURST READY) + edge glow + nameplate borders already convey everything actionable.
• Demo chat spam silenced. /acc test (and /acc test bg, /acc test world) previously printed the beat-by-beat note to chat — 7 lines for the arena demo. Now only the start + end banners fire by default. The per-beat notes still print when verbose mode is on.

Added

• /acc verbose [on|off] — new slash command that toggles db.frame.verbose. When on, the HUD reverts to the v2.3.1 information density (full callout list, comp badge, chain title + step lines) and the demo chat spam returns. When off (default), you get the Quiet HUD. Persists across /reload. Aliases: none (it's an additive toggle, not a master switch).
• New db.frame.verbose SavedVariable key (defaults to false). Existing installs auto-merge it on next login via the standard DEFAULTS merger; no migration needed.

Notes

• This is a UI-render-only change. The engine still emits the full callout list, comp identification, and chain data on every evaluation — /acc trace dump, the bridge API (_G.ArenaCoachTBC.GetCallouts(), GetChain(), GetEnemyCompLabel()), and WeakAura consumers see everything unchanged.
• Tests: 608 still passing. Updated three demo specs in Tests/Core_spec.lua to set db.frame.verbose = true before running so the per-beat chat assertions still hold.

[2.3.1] - 2026-05-25

Fixed

• Recommendation frame leaked internal score-contributor identifiers as text ("PRIEST [role_healer(25), trinket_down(20), health_below_50(30)] | RMP_DISC_3V3 spec-confirmed (1.00)"). User report: "there are random words like lkjasfsa_lajfda, seems like not properly translated or mapped to proper spells." Root cause: UI:Apply was rendering recommendation.reason verbatim for KILL/SWAP/OPEN modes — but that field is dev-only, meant for /acc trace dump, never user-facing. The mode label + target name + target stats row (HP / kill prob / BURST READY) + callouts list + comp badge + chain block already carry everything the user needs. Now UI:Apply renders reason only via the localised reasonKey path (DEFEND / RESET) and drops the raw debug text otherwise.
• Chain step lines were redundantly tagged with their category ("Step 1. Sap (INCAPACITATE)", "Step 2. Polymorph (INCAPACITATE)", "Step 3. Kidney Shot (STUN)"). The category is already implicit in the chain title; the parenthetical was visual noise. Dropped — now just "Step 1. Sap", "Step 2. Polymorph", "Step 3. Kidney Shot".

Notes

• The recommendation.reason field is unchanged — /acc trace dump, the bug report, and any WeakAura consumer that reads _G.ArenaCoachTBC.GetReason() still get the full debug breakdown. This is a UI-render-only change.
• Tests still at 608 passing, locale parity at 110/110.

[2.3.0] - 2026-05-25

Quality release. One real bug fix on top of v2.2.6, plus a sweep of dead code and stale docs that had piled up across the v2.1-v2.2 patch cycle.

Fixed

• /acc test (and any non-arena trigger of UI:Apply) only painted the text frame — no screen edge glow, no nameplate highlight. User report: "I do not see any HUD in the latest release, only text." Root cause: UI:Apply's v2.2.0 visual-layer block gated on inPvP = (arena|bg|world) and ignored the _forceShow flag the demo sets to bypass the v2.2.5 auto-hide. Result: the early hide gate let the rec through, but the later edge-glow + nameplate gate still required real PvP context and so nothing painted on the periphery. Fixed with a 3-line change: local showVisualLayers = inPvP or forceShow. Regression-tested in Tests/UI_spec.lua (test #607 / #608).

Removed

• UI.lua dead code (38 lines): makeIcon(parent, size) (lines 27-59) and spellIcon(spellID) (lines 62-67) — orphaned since v2.2.1 removed the icon rows that called them. Module header line about "two icon rows" updated to reflect the v2.2.0 visual-layer architecture.

Changed

• Docs refresh — first comprehensive sweep since v2.0 shipped 9 patches ago:
  • docs/weakaura-pack.md: removed the Path 1 paste-string section that directed users to a tool we deleted in v2.2.6. Added a deprecation note explaining the parser-library limitation. Path 2 (trigger-code snippets for hand-built WAs) unchanged.
  • docs/architecture.md: title bump v2.0 → v2.2; new sections for ScreenEdgeGlow.lua, Nameplate.lua, the v2.2.5 auto-hide gate + /acc off master switch; Sounds.lua description corrected (numeric SoundKit IDs, not the broken .ogg paths from before v2.1.6).
  • docs/manual-smoke.md: slash-command checklist extended with /acc off, /acc on, /acc glow, /acc nameplate; new HUD smoke step that asserts the full visual stack paints during /acc test; new city-lag smoke step; added Plater / KuiNameplates / TidyPlates to the addon-conflict matrix.
  • ArenaCoachTBC/README.md: stale Interface: 20504 reference bumped to 20505; slash-command table expanded with the v2.2 commands.
• Roadmap refresh — first sweep since v2.0:
  • ROADMAP.md (v1): marked the items shipped through v2.2.6 as done with their actual shipping vehicle. Crossed-out the "Prepackaged WeakAura export string" item with a link to the v2.2.6 abandonment note. Moved M6 hardening items (debugprofilestop assertions, memory fuzz, multi-Lua CI matrix, headless replay tool, evaluation server) into a "Deferred — future hardening" section. Marked the M5 cloud-telemetry item as deferred indefinitely.
  • ROADMAP-v2.md: added a "What shipped after M12" section that one-line-summarises every v2.1, v2.2, v2.3 release with a CHANGELOG anchor, so future readers understand why v2.0's "complete" still got 9 patches stacked on top.

Notes

• Test count 606 → 608 (+2 regression tests for the HUD-demo bug; the v2.1.3 tests that briefly broke during development now pass cleanly).
• Locale parity green: still 110 keys per locale (enUS, zhCN). No new locale work in this release.
• No behaviour changes outside the bug fix. No new slash commands, no new SavedVariables keys, no schema migrations.

[2.2.6] - 2026-05-25

Removed

• WeakAura paste-string export pipeline. v2.0–v2.2.5 shipped pre-built !WA:2! import strings in docs/weakaura-imports.md + README.md, generated by tools/export_weakauras.mjs via the node-weakauras-parser npm package. After 6 patches chasing import failures (parser format, internalVersion, version=3, semver, config/information shape) the root cause turned out to be the parser itself: even re-encoding a known-working Wago WA byte-for-byte produces a string that decodes correctly but fails WA's import-validator byte check (no Import button shown). Removed the whole pipeline:
  • tools/export_weakauras.mjs (the broken generator)
  • tools/package.json, tools/package-lock.json, tools/node_modules/ (npm deps)
  • docs/weakaura-imports.md (auto-generated output)
  • docs/wa-hello-test.md + tools/test_hello_wa.mjs (diagnostics from the chase)
  • The "Paste-ready import strings" section in README.md (replaced with a note explaining the limitation + pointing at docs/weakaura-pack.md for trigger source code users can hand-build a WA from)
• The _G.ArenaCoachTBC bridge API and WeakAuras.ScanEvents("ACC_RECOMMENDATION", rec) event remain unchanged — power users who want custom auras still have everything they need.

Why

The addon's built-in HUD (v2.1.6 + v2.2.0) already renders the mode badge, target stats, screen edge glow, nameplate highlight, and audio cues. WA paste-strings were redundant convenience for users who wanted the same display in their own UI framework — not worth shipping a broken pipeline to chase.

[2.2.5] - 2026-05-25

Fixed

• Major frame-rate drop in cities on PvP-flagged characters. User report: addon caused lag in main city. Root cause: onNameplateChange re-ran Evaluate() whenever the player was PvP-flagged (world_idle context) and any nameplate appeared / disappeared — which in Stormwind is hundreds of events per second. Engine had nothing to recommend (no hostile contact), so the work was pure waste. Gate now only triggers Evaluate when pvpContext == "bg" or "world" (an actual fight). world_idle (flagged + no enemies) no longer drives evaluation.
• Frame stayed visible with stale rec outside PvP. Previously after /acc test or after leaving an arena, the recommendation frame would linger center-screen showing the last computed rec. Now UI:Apply checks Core.state.pvpContext and hides the frame + edge glow + nameplate paint when the context is explicitly "none" (no PvP relevance) or "world_idle". The /acc test demo bypasses this via a per-beat _forceShow flag so the walk-through still renders end-to-end.

Added

• /acc off / /acc on master switch (aliases: /acc disable / /acc enable). Sets db.enabled and immediately hides every visual layer. Persists across /reload and login sessions. Same enabled flag the engine already short-circuits on, so no Evaluate work happens while off.

[2.2.4] - 2026-05-25

Fixed

• WeakAura imports still failed in v2.2.3 — the real root cause. Field-by-field deep-diff against a known-working Wago WA revealed that d.config and d.information were declared as empty objects ({}) in the exporter but the working Wago WA had them as empty arrays ([]). The parser encodes the two shapes differently — {} becomes a Lua hashmap, [] becomes a Lua sequence — and modern WA's import validator rejects the hashmap shape so hard that the import preview dialog never even surfaces (user report: "no Import button appears after pasting"). Changed both to [] in tools/export_weakauras.mjs. Verified post-regeneration: the only remaining schema differences against the Wago reference are field ordering (which doesn't matter for Lua tables) plus the expected per-aura fields (uid, url, semver, wagoID).

[2.2.3] - 2026-05-25

Fixed

• WeakAura import dialog never showed the Import button. Decoded a known-working Wago WA (v1TwWSgUh — a user-uploaded copy of our Mode badge) side-by-side with our generated string. Three schema gaps emerged:

  • d.version was 1 — modern WA expects 3. Version 1 is so old that newer WA builds silently drop the import without surfacing the dialog (no Import button shown).
  • d.semver was undefined — required for version-3 imports as a user-visible release tag.
  • d.internalVersion (fixed in v2.2.2) was already set to 90, so that part was correct.

  Set version: 3 and semver: '2.2.3' in the exporter COMMON block. All 5 templates regenerated in docs/weakaura-imports.md and the inline copies in README.md. Now matches the schema of the proven-working Wago WA byte-for-byte (modulo the per-aura uid).

[2.2.2] - 2026-05-25

Fixed

• WeakAura imports showed "this aura was created with a very old version" warning. Our generated d object had internalVersion undefined, which WA-Classic reads as "ancient" and triggers schema-migration warnings (or outright rejects on stricter builds). Set internalVersion: 90 in the exporter COMMON block — matches current upstream WeakAuras (latest WeakAuras/WeakAuras.lua on main defines local internalVersion = 90). All 5 templates regenerated in docs/weakaura-imports.md + the inline copies in README.md. Imports cleanly without the version warning.
• Also corrected the WA url field from the old wow_tbc_arena_pvp_strategy repo slug to the renamed ArenaCoachTBC slug, so the "Source" link on the WA matches the current GitHub URL.

[2.2.1] - 2026-05-25

Fixed

• WeakAura import strings rejected by WA-Classic. v2.1.6 switched the exporter to FormatVersion 1 (!-prefixed Deflate) on a wrong hunch — actual round-trip verification through Wago showed WA-Classic accepts the FormatVersion 2 (!WA:2! binary-serialized) format that the exporter had always used. Reverted tools/export_weakauras.mjs to FormatVersion 2 and regenerated all 5 templates in docs/weakaura-imports.md + the inline copies in README.md. The strings now import cleanly.
• Dead icon rows at the bottom of the recommendation frame. Since v1 the frame rendered 14 friendly + 9 enemy cooldown reminder icons across two rows at the bottom, but UI:UpdateIcons was only ever called from tests — no production code fed it a "ready set", so every icon sat at its initial 0.4 alpha forever and communicated nothing. Stripped the icon rows, the populate function, the unused UpdateIcons method, and the frame.compactMode toggle that gated their visibility. Frame height drops from 170px to 110px so the HUD is more compact. UI_FRIENDLY_CDS / UI_ENEMY_CDS locale keys removed from both locales (110 keys each, parity green).

Removed

• UI.friendlyIcons, UI.enemyIcons, UI:_PopulateIconRows, UI:UpdateIcons — dead UI code.
• db.frame.compactMode SavedVariable — there's nothing to toggle now that the icon rows are gone. Existing installs that have this set in SavedVariables will see it simply ignored on next login (no migration needed).
• 7 tests that exercised the removed surface (Apply compactMode show/hide, UpdateIcons happy/nil paths, icon button spellID, icon tooltip OnEnter/OnLeave, _PopulateIconRows fallback). Test count 613 → 606.

[2.2.0] - 2026-05-25

Eyes Up. Two new peripheral-vision layers on top of the v2.1.6 HUD, so the engine's call reaches you even when your eyes are on the action — not on the frame.

Added

• Mode-coloured screen edge glow. A pulsing band hugs the four screen edges, coloured to match the active recommendation (red KILL, orange SWAP, blue DEFEND, yellow OPEN). Pulse period 1.6s, alpha breathes between 0.18 and 0.42 so it stays visible but never dominates the viewport. RESET intentionally has no colour — between fights the glow goes dark instead of strobing for nothing. New ArenaCoachTBC/ScreenEdgeGlow.lua module. Toggle: /acc glow on|off (default on). Gated by PvP context — only renders in arena / BG / world PvP, never in idle world.
• Nameplate highlight for kill / swap targets. The engine's primary target gets a red border on its nameplate; the swap candidate (when in SWAP mode) gets an orange border. Adds a child overlay frame to each affected nameplate so we coexist cleanly with Plater / KuiNameplates / TidyPlates (we never modify the native health bar / cast bar / name text). New ArenaCoachTBC/Nameplate.lua module. Toggle: /acc nameplate on|off (default on). Hook driven by the existing NAME_PLATE_UNIT_ADDED / REMOVED subscriptions; per-Apply ClearAll + reapply keeps state coherent through plate cycling.
• Two new slash commands: /acc glow [on|off] and /acc nameplate [on|off], with bare-toggle behaviour when no argument is passed.
• db.alerts.edgeGlow and db.alerts.nameplate SavedVariable keys, both default true. Existing installs auto-merge on next login via the standard DEFAULTS merger.
• 12 new tests (ScreenEdgeGlow_spec.lua: 6 cases for colour table + SetMode/Hide round-trip; Nameplate_spec.lua: 6 cases for Apply / Highlight / ClearAll idempotence + overlay lifecycle). Test count 601 → 613.

Notes

• Both new visual layers are arena/BG/world-gated — in idle world (no hostile context) the glow + nameplate paint are skipped to avoid being a constant visual distraction.
• The base frame's bigText (v2.1.6) + audio cues (v2.1.6) + edge glow (v2.2.0) + nameplate (v2.2.0) together form the "Eyes Up" feature pack. Each layer is independently toggleable so users can pick the subset that doesn't conflict with their existing UI.

[2.1.6] - 2026-05-25

Fixed

• Audio cues were silently broken since v1.0. Sounds.lua referenced Sound/Voice/*.ogg paths that were never bundled in the addon zip, so every PlaySoundFile invocation failed and "audio callouts" did nothing in any release. v2.1.6 rewires Sounds:Play and the new Sounds:PlayMode to numeric TBC Classic SoundKit IDs (RaidWarning chime, RaidBossEmote alert, PvPVictory chord, queue ding, quest pop) that ship with the WoW client itself. No new assets needed; cues fire reliably on every install. The db.alerts.sound toggle works as advertised.
• Mode-transition audio. Pre-v2.1.6 the only audio cue was the per-callout sound, which fired on CALL_HOJ_KILL / CALL_TREMOR_FEAR / BURST_NOW events. v2.1.6 adds Sounds:PlayMode(mode) driven from UI:Apply that plays a distinct ding when the recommended mode flips (KILL / SWAP / DEFEND / OPEN), so even with your eyes off the frame you hear the engine's call. Same alerts.sound gate; arena-only for the same noise-floor reason.

Added

• Bigger, more readable mode label. f.bigText upgraded from GameFontNormalHuge (~22pt) to a custom 32pt outlined font. The mode + target line is now legible from across a battleground screen, not buried in the corner.
• Target stats row. A new f.statsText line below the mode label renders HP <n>% kill <n>% BURST READY when the rec carries a primary target. Hidden on DEFEND / RESET (no target). Engine now emits primaryTargetHp (0..1) and killProb (0..1) on the recommendation table so the HUD has the data it needs without reaching into state.
• 3 new locale keys (109 → 112 per locale, parity green): UI_HP_LABEL, UI_KILL_PROB_LABEL, UI_BURST_READY.

Notes

• This is the first half of v2.2 "Eyes Up" (visual + audio overhaul). The remaining items — mode-coloured screen edge glow, nameplate highlight for the kill / swap targets — will land as v2.2.0 once the HUD changes have been validated in real combat.

[2.1.5] - 2026-05-25

Added

• CurseForge auto-upload wired up. Added ## X-Curse-Project-ID: 1552792 to the TOC. The release workflow's BigWigs packager step (already fixed in v2.1.4) now has both the API token (CF_API_KEY GitHub secret) and the project ID needed to publish each tagged release straight to the CurseForge ArenaCoachTBC project page. Tagged releases (vX.Y.Z) auto-push the addon zip to CurseForge; dev prereleases (vX.Y.Z-dev.N from main pushes) still only publish to GitHub.

Notes

• Wago upload still pending — WAGO_API_TOKEN secret is set but no ## X-Wago-ID: in the TOC yet. Add the Wago slug to the TOC + cut another patch to enable Wago uploads.

[2.1.4] - 2026-05-25

Added

• TBC Anniversary client support. ## Interface: bumped from 20504 (BCC 2.5.4) to 20505 (Anniversary 2.5.5). Same Interface-BCC line keeps Burning Crusade Classic clients working — addon now loads cleanly on both the closed BCC era and the live Anniversary realms without an "out of date" warning. Title / Notes updated to advertise "Anniversary / Classic PvP — arena, BG, world" instead of the older "TBC Classic arena" framing.

Fixed

• Release workflow's BigWigs packager step no longer fails with Could not find an addon TOC file. Cause: the addon TOC lives at ArenaCoachTBC/ArenaCoachTBC.toc but the packager defaults to looking for the TOC at the repo root. Passing args: -t ArenaCoachTBC to the action's release.sh sets the project topdir to the addon subdir, so the packager finds the TOC, reads the in-addon .pkgmeta (ArenaCoachTBC/.pkgmeta, new), and produces a clean zip. This unblocks CurseForge + Wago uploads once project IDs are wired up.
• New ArenaCoachTBC/.pkgmeta declares package-as: ArenaCoachTBC, enable-nolib-creation: no, and ignores Tests/ + .luacheckrc. The root .pkgmeta is now dead code (left in place to avoid breaking external tooling; will be removed in a later cut).

Notes

• The CurseForge + Wago API tokens are configured as GitHub secrets (CF_API_KEY, WAGO_API_TOKEN), but project IDs are still required in the TOC (## X-Curse-Project-ID: / ## X-Wago-ID:) before uploads can succeed. Create the CurseForge + Wago projects for ArenaCoachTBC and add the IDs to ArenaCoachTBC.toc to complete the publishing chain. Without IDs the packager will skip the upload step (the GitHub Release will still publish).

[2.1.3] - 2026-05-25

Fixed

• DEFEND / RESET modes no longer show a target name. Reported via a WSG screenshot: the frame displayed "DEFEND: lhealyoupeel" (in Chinese: "守: lhealyoupeel"). Reading that as "defend against lhealyoupeel" is the opposite of the intent — DEFEND is about your team's defensive cooldowns, not a target to attack. UI:Apply now restricts the "<mode>: <name>" form to OPEN / KILL / SWAP only; DEFEND and RESET render mode alone.
• DEFEND reason text now follows the WoW client locale. Same screenshot showed "defensive: trained" in English next to Chinese callouts. Engine now emits rec.reasonKey (e.g. REASON_DEFEND_TRAINED) for the six known DEFEND reasons + the RESET case; UI prefers reasonKey through L() over the raw debug reason string. Chinese client now sees "防御 - 治疗被集火". KILL / SWAP / OPEN reasons stay as the raw English contributor-list text (they carry variable per-evaluation data, not a stable key).
• 7 new locale keys (109 per locale, parity green): REASON_DEFEND_TRAINED, REASON_DEFEND_LOW_HEALER, REASON_DEFEND_ENEMY_LUST, REASON_DEFEND_MULTI_BURST, REASON_DEFEND_HEALER_CC, REASON_DEFEND_TRIPLE_DPS, REASON_RESET.
• 8 new tests (3 in UI_spec for the target-suppression cases + reasonKey rendering, 4 in StrategyEngine_extra_spec for the reasonKey field on DEFEND / RESET / KILL).
• Mock harness mockMethods:SetText/GetText added so UI specs can assert on rendered text.

[2.1.2] - 2026-05-25

Fixed

• Frame stayed at "Awaiting opener..." in WSG / BGs / world PvP when no combat was happening. Reported: "doesn't work in WSG". Root cause: Core:Evaluate runs the non-arena enemy refresh (RefreshEnemiesNonArena), but Evaluate itself only fired on arena events / CLEU / aura events. While running across a BG map with no combat, no event ticked, so the engine never re-scanned nameplates and never saw enemies become visible. v2.1.2 subscribes to NAME_PLATE_UNIT_ADDED / NAME_PLATE_UNIT_REMOVED and re-evaluates whenever the player's nameplate set changes — only when pvpContext == "bg" / "world" / "world_idle" (arena keeps the original event-driven flow). Headless tests cover BG context evaluation, non-PvP context ignoring nameplate events, and the world-context re-evaluation path.

You do not need a WeakAura to make ArenaCoachTBC work in BG. The addon's own frame should populate naturally as you run past enemies. The WA bridge (including _G.ArenaCoachTBC.GetPvPContext() from v2.1.1) is still there if you want a custom HUD on top, but the default frame is the supported path.

[2.1.1] - 2026-05-25

Polish + visibility on top of v2.1. Same engine; surface and docs upgraded.

Added

• /acc test bg — 5-beat BG walk-through (engaged → flag carrier picks up → flag carrier low HP → CALL_BG_DEFEND on train → reset). Walks the BG scoring branches the same way /acc test walks arena.
• /acc test world — 4-beat world PvP walk-through (engaged → push burst → DEFEND on low HP → reset). Demonstrates the single-target focus / no-SWAP-thrash behaviour.
• WeakAuraBridge.GetPvPContext() exposes state.pvpContext to WeakAuras so consumers can render different displays per context (e.g., hide the comp badge in BG, show flag-carrier-specific text in world).
• AV-scale perf test (Tests/Performance_spec.lua) — 40-enemy state, asserts SE:Evaluate stays under 50ms CI budget. Confirms the v2.1 engine scales to AV without code changes.

Changed

• README "Works in every PvP context" matrix added at the top — bilingual, summarises arena / BG / world / duel behaviour at a glance.
• README slash-command table now lists /acc test bg and /acc test world (bilingual).
• Core:_RunTestDemoMode(beats, label) refactored to accept the beat list + label as arguments (was hard-coded to the arena RMP beats); /acc test dispatches into one of three beat sets.

Tests

586 → 590 (+4). New tests: /acc test bg runs and prints the BG walk-through banner; /acc test world does the same for world; WAB:GetPvPContext round-trips state; AV 40-enemy Evaluate stays within perf budget.

[2.1.0] - 2026-05-25

Wild PvP — battlegrounds + world PvP + duels. The addon used to be effectively disabled outside arena because enemy discovery was hardcoded to arena1..arena5 unit IDs and bracket-aware scoring assumed 2/3/5 teams. v2.1 extends the engine to BG (WSG / AB / AV / EotS) and open-world PvP without breaking the arena flows that already work.

User-facing pitch: open the addon in a Warsong Gulch queue and you'll see a recommendation frame with sensible kill targets (the flag carrier at low HP dominates), low-HP straggler swaps, and BG-flavoured callouts. Toggle nameplates and enemies appear / disappear as they come into LOS. Duels light up the frame with your opponent. None of this required new modules — same engine, just made context-aware.

12 new locale keys (103 → 115). 50+ new tests (538 → 586+). CI 99% coverage gate still green.

Added — M16 (v2.1 quality + ship)

• End-to-end BG simulation. Tests/BGModeE2E_spec.lua synthesises a 10-player BG roster and verifies: flag-carrier dominates kill priority, low-HP straggler swap, CALL_FLAG_CARRIER_LOW + CALL_BG_DEFEND emission, no comp identification, no SWAP thrash on small score gaps, perf within budget (10 enemies × Evaluate stays under 30ms CI), arena-only callouts don't fire spuriously. 8 cases.
• Per-source pattern progress (bug fix). Patterns:Observe(spellID, ts, sourceGUID) now keys progress by <patternId>|<sourceGUID> so two enemy priests casting Psychic Scream don't collide / false-complete each other's chains. Legacy 2-arg signature (spellID, ts) still works (sourceGUID defaults to a sentinel). Probability accepts an optional sourceGUID for per-caster lookup; without it, returns MAX progress across tracked sources. 4 new tests.
• Version bump: 2.0.2 → 2.1.0 (minor — new feature surface).

Added — M15 (v2.1 world PvP + duels)

• World PvP engine branch. When state.pvpContext == "world":
  • decideMode skips OPEN (no arena planning phase) and skips SWAP (single-target focus, no team coordination)
  • Strategies:Identify is bypassed (matched comp would be coincidence in a fixed-roster-less context)
  • shouldDefend's comp-based triple_dps_pre check is bypassed for the same reason
  • DEFEND still fires when the player's HP drops below the aggression-tuned threshold (via existing lowestHealer path — in world the "lowest friendly healer" is just the player themselves)
• BG mode also skips OPEN (same reasoning — no pre-combat planning in BG).
• Duel detection. New event handlers in Core:
  • DUEL_REQUESTED → forces pvpContext = "world", stamps Core._lastWorldHostileTs, seeds an enemy entry from the current target via _NonArenaCLEUStub + refreshUnit. Triggers an immediate Evaluate so the frame populates as soon as the duel countdown starts.
  • DUEL_FINISHED → clears the recent-hostile timestamp and re-runs DetectPvPContext to drop back to whatever context the player is in.
• 8 new tests (6 in StrategyEngine_extra_spec, 2 in Core_spec): world PRE → no OPEN, world SWAP suppression, world skips comp ID, arena comp ID regression, world DEFEND on low HP, BG PRE → no OPEN, duel start populates target, duel end clears + re-detects.

Added — M14 (v2.1 BG mode)

• BG scoring boosts. Three new SE.weights entries active when state.pvpContext == "bg":
  • bg_flag_carrier = 200 — WSG flag aura (23333 Alliance / 23335 Horde) eclipses every other priority
  • bg_low_hp_straggler = 30 — bonus for any enemy <30% HP (BG produces lots of swap windows)
  • bg_healer_boost = 10 — small bump on top of role_healer (healer death decides BG fights)
• BG SWAP threshold tightened to 30 (vs default 10) to prevent thrash in messy BG combat where LOS and target reshuffle.
• BG callouts. buildCallouts emits when pvpContext == "bg":
  • CALL_FLAG_CARRIER_LOW when the kill target has a flag aura + <50% HP
  • CALL_BG_DEFEND on DEFEND mode (cleaner cue than the arena-flavoured Pain Sup / BoP set)
• 5 new locale keys, parity green at 103 each: CALL_FLAG_CARRIER_LOW, CALL_INCOMING_PLAYERS, CALL_BASE_UNDER_ATTACK, CALL_BG_DEFEND, CALL_BG_RES_TIMER. (Three are wired in this PR; two are reserved for future BG-objective work in v2.2.)
• Class-prior tier in OpponentProfile. PUG'd BGs reset team-signature profiles every match. New OP:GetClassPrior(class, db) / OP:UpdateClass(class, key, observed, db) track tendencies across all observations of a class regardless of team. OP:EstimateWithClassPrior(profile, key, class, default, db) prefers the team profile when it has ≥5 samples, falls back to the class prior, falls back to the default. Stored under db.classPriors[CLASS][tendency] — does NOT mix with arena's db.profiles.
• 13 new tests (6 in StrategyEngine_extra_spec, 7 in OpponentProfile_spec).

Added — M13 (v2.1 foundation)

• PvP context detection. New Core:DetectPvPContext() returns one of "arena" / "bg" / "world" / "world_idle" / "none". Reads IsActiveBattlefieldArena(), GetInstanceInfo(), UnitIsPVP("player"). Cached on state.pvpContext. Refreshed by PLAYER_ENTERING_WORLD, ARENA_OPPONENT_UPDATE, and the new ZONE_CHANGED_NEW_AREA subscription. Headless-permissive — when WoW APIs are absent, the cached fixture value survives.
• Non-arena enemy discovery. New Core:RefreshEnemiesNonArena() walks nameplate1..nameplate40, keeps hostile players, keys by GUID (not unit ID, since nameplate units reshuffle on LOS). Core:_NonArenaCLEUStub(guid, name) creates a stub entry from CLEU events when the nameplate isn't visible yet. 30-second TTL prunes entries we haven't re-observed. Arena entries (keyed by arenaN) are left alone — RefreshArenaEnemies retains ownership.
• Core:Evaluate routes between RefreshArenaEnemies (arena context) and RefreshEnemiesNonArena (bg/world). No-op outside PvP.
• UI:Apply gate updated — prefers state.pvpContext == "arena" over the v2.0.2 IsActiveBattlefieldArena() direct call; falls back to the API if Core hasn't populated state yet (early-load).
• Core:UpdateRating early-returns when pvpContext ≠ "arena". Avoids the WoW API roundtrip in BG/world and prevents bracket=10 (WSG team size) from accidentally indexing into the rated-info table.

15 new tests in Core_spec covering each context, nameplate discovery, CLEU stub creation, stub non-overwrite, TTL prune, arena-key preservation, rating-API gate. 553/553 green.

[2.0.2] - 2026-05-25

Bilingual docs + WSG/BG flash gate.

Fixed

• Screen flash + voice cues no longer fire outside arena. Reported: "it kept blinking redness when I am doing WSG". The URGENT-mode screen flash + voice callout dispatch were unconditional once db.alerts.screenFlash / db.alerts.sound were on. In BG (and world PvP / outside-PvP) the engine's DEFEND-trigger heuristics fire spuriously and pulsed red flash every few seconds. UI:Apply now gates both the screen flash and the voice cue on IsActiveBattlefieldArena() — the recommendation frame itself stays available (so you can read the data in BG), but the intrusive alerts only fire in actual arena instances. Headless tests pass the gate via the missing-API permissive branch. +2 new tests in UI_spec.
• Tooltip locale carry-over fix from v2.0.1 continues to work (SetSpellByID + GetSpellInfo fallback chain).

Changed

• All user-facing docs are now bilingual (English + 中文) inline. No separate language files. The root README.md, ArenaCoachTBC/README.md, docs/architecture.md, docs/weakaura-pack.md, and docs/weakaura-imports.md now interleave English and Chinese content at the section level. Section headings carry both languages (Installation / 安装); prose paragraphs appear in both languages back-to-back. Code blocks, slash commands, API names, and config keys are kept English (those are universal artifacts). tools/export_weakauras.mjs updated to emit bilingual headers + descriptions, so future regenerations keep both languages.

Documented

• BG support is partial in v2.0.2. The recommendation frame can be shown in BG via /acc toggle, and the engine continues to evaluate (so WeakAuras and the trace log still receive data), but the per-comp catalog and chain definitions are tuned for arena 2v2/3v3/5v5. Proper BG support (large-team mode, BG-specific kill priority, BG-specific chains) is a v2.1 roadmap item.

[2.0.1] - 2026-05-25

Documentation + UX polish on top of v2.0. No engine changes — 536 tests, 99%+ coverage, 81% benchmark baseline.

Fixed

• Mouse-over tooltips on icon-row buttons now follow the WoW client locale instead of showing hardcoded English. makeIcon's OnEnter calls GameTooltip:SetSpellByID(spellID) (canonical localized tooltip with icon + name + flavor text) when available, with GetSpellInfo(spellID) as a localized fallback. The English string label is kept only as the last resort when both WoW APIs are absent (headless tests). Icon buttons now carry spellID directly for this path.

Changed

• /acc test is now a DBM-style scripted UI walk-through instead of a tight chat-only loop. Force-shows the frame, then steps through 7 beats over ~14 seconds via C_Timer.After: OPEN → KILL → BURST_NOW pulse → SWAP → DEFEND (with screen flash if db.alerts.screenFlash is on) → profile-driven callout (CALL_SAVE_TREMOR_HOJ) → RESET. Each beat re-uses the real UI:Apply + WeakAuraBridge:Publish path so voice cues, chain block, comp badge, and burst pulse fire exactly as they would in a real arena. Frame visibility is saved + restored. Legacy chat-only summary kept under /acc test print (still walks Strategies.testComps).

Added

• Programmatic WeakAura import strings. tools/export_weakauras.mjs (node-weakauras-parser) generates 5 paste-ready !WA:2!... strings to docs/weakaura-imports.md. Round-trip validated — each string decodes back to a valid WA config table. Templates: Mode badge, Burst gate, Defensive alert, Callout stream, Comp readout. Re-runnable: cd tools && npm install && node export_weakauras.mjs.
• docs/weakaura-imports.md — the generated paste-ready strings, ready for /wa → Import.
• _design/ArenaCoachTBC Design Showcase.html — 9-scene scrollable HTML mood board documenting the v2.0 user experience: hero, anatomy of the frame, in-arena KILL / DEFEND scenes, chain anatomy close-up, opponent profile in action, settings panel, WeakAura bridge integration, compact vs full mode comparison. Dark tactical-HUD aesthetic; mode colours pulled verbatim from UI.lua > modeColors; no Blizzard IP.

Docs

• README rewrite — full step-by-step installation (per-OS paths), first-run checklist, daily-usage walkthrough during arena, complete slash-command reference table, configuration knob reference, localisation note clarifying that spell IDs are universal and names come from GetSpellInfo(spellID) in the WoW client's locale.
• Removed the stale "tuned for 5v5 melee cleave" framing. v2 adapts to any composition — OwnComps:Infer + 5 archetypes (MELEE_CLEAVE, CASTER_CLEAVE, DRAIN, JUNGLE, DOUBLE_HEALER), explained inline.
• docs/weakaura-pack.md restructured around two paths: paste-ready import strings (Path 1, recommended) vs hand-built trigger code (Path 2, DIY).
• ArenaCoachTBC/README.md slash-command table updated to reflect new /acc test behaviour.

Locale

• TEST_DEMO_START, TEST_DEMO_END, TEST_DEMO_NO_UI added to enUS + zhCN (98 keys per locale, parity gate green).

Tests

• Three new tests in Core_spec.lua: RunTestMode (default) emits start + per-beat + end lines (≥8); RunTestMode "print" triggers the legacy summary; demo restores hidden frame state when it started hidden. 533 → 535 total.

[2.0.0] - 2026-05-25

v2.0 ships the engine-depth roadmap: spec-aware comp matching, multi-link CC chain planning, per-opponent Bayesian profiles, multi-step lookahead with bounded branching, pattern recognition for recurring kill setups, rating-aware risk gating, and a calibrated kill-probability model. 195 new tests (350 → 545), 89 → 95 locale keys per locale, 5 new pure modules (Chain, OpponentProfile, Lookahead, Patterns, Sounds), and a benchmark suite reporting 81% baseline agreement against hand-labelled scenarios.

The user-visible pitch: your coach now learns your opponents. A team that always trinkets Fear stops getting the "tremor for fear" callout — Tremor gets saved for HoJ instead. A mage that consistently Ice Blocks at 30% HP causes burst to be held. None of this is hardcoded — it learns from /acc record logs, per-team, no names persisted.

Fixed

• Live arena observations now scan unit auras for Mortal Strike, Windfury, Bloodlust/Heroism, and enemy burst buffs before each evaluation, so burst gating is driven by observed state instead of permanently missing msActiveOn / windfuryActive flags.
• Train detection now counts damage events only when they land on friendly healers, preventing ordinary melee damage on DPS teammates from forcing DEFEND mode.
• Arena unit refresh now clears stale class/GUID/spec data when an arenaN or party unit disappears, preventing old rosters from polluting later comp identification.
• EventBus handler failures are captured by ErrorReporter, so /acc bugreport includes real in-addon handler errors instead of only manually captured failures.
• Cold Snap and Icy Veins now use distinct spell IDs and cooldown durations.
• Dev prerelease tags (vX.Y.Z-dev.N) are skipped by the release workflow's tag-trigger path so they cannot be republished as stable releases.

Changed

• Strategy scoring now applies openTarget / swapTarget hints from the comp catalog, exposes secondaryTargetClass, and makes strategy.aggression affect the SWAP threshold.
• CI now runs the standalone StrategyEngine_spec.lua smoke spec in addition to the coverage suite.

Added

• Profile-driven callouts (#65). buildCallouts consults OpponentProfile:EstimateOrDefault for three binary tendencies and emits the matching callout when the posterior mean ≥ 0.7 and the sample is opinionated (n ≥ 5): CALL_FAKE_KICK_2 ("they kick the first heal — fake your second") gated on kicksFirstHeal; CALL_SAVE_TREMOR_HOJ ("they trinket Fear — save Tremor for HoJ") gated on trinketsFear; CALL_BURST_BLOCK_INCOMING ("Ice Block expected — hold burst") gated on iceBlockBelow30. Core.PrepareAndEvaluate resolves the opponent profile via OpponentProfile:Signature(state.enemies) and attaches it to state.opponentProfile (+ state.opponentSignature) before calling StrategyEngine.Evaluate — the engine itself remains pure. New rec.profileContrib field captures the comma-joined <tendency>=<mean> pairs that contributed; the /acc trace snapshot extends with a profileContrib field for post-mortem inspection. rec.opponentSignature exposed via WAB pipeline. 6 new tests cover each tendency gate, the threshold suppression at n<5, no-profile suppression, and the signature pass-through. Three new locale keys per locale (89 each).
• Bayesian update variants + Estimate with CI + fallback (#64). OP:UpdateBinary(profile, key, observed) mirrors OP:Update but operates on an already-resolved profile reference (no signature/db lookup). OP:Estimate(profile, key) returns { mean, low, high, n } with a normal-approximation 95% confidence interval on the Beta(α, β) prior. OP:EstimateOrDefault(profile, key, compDefault) returns compDefault when n < OP.MIN_SAMPLES_FOR_OPINION (default 5), else the posterior mean — the gating primitive #65's profile-driven callouts will use. Tests verify convergence (20 positive observations → mean > 0.85), CI shrinks with sample size, and the fallback gate engages below threshold.
• OpponentProfile module (#63, opens M9 — the keystone). New OpponentProfile.lua stores per-opponent-team behavioural profiles keyed by team signature (<sorted_classes>#<djb2_hash_of_sorted_names>). Four binary tendencies tracked as Beta(α, β) priors: trinketsFear, iceBlockBelow30, kicksFirstHeal, sapsPriest. API: Signature(enemies), Get(sig, db), Update(sig, event, db), Forget(sig, db), Mean(profile, tendency), SampleCount(profile, tendency). Update({tendency, observed}) bumps α (observed=true) or β (observed=false). Get backfills any newly-added tendency in the canonical list onto older persisted profiles. Names are NEVER stored — the djb2 hash is the only identifier that survives to SavedVariables (per the v2 "no personally-identifying data persists" rule). New SavedVar db.profiles = {}. Pure module: never touches a WoW API; reads / writes a passed-in db. WeakAuraBridge exposes GetOpponentProfile(), GetOpponentSignature(), and GetTendencyMean(tendency) for WeakAuras to read the current opponent's profile without re-implementing the signature logic. 23 new tests (21 in OpponentProfile_spec, 2 in WAB_spec) covering signature determinism, class-set + name sensitivity, name non-leakage in stored shape, Beta update math, mean / sample-count, backward-compat backfill, forget, and the WAB plumbing. M10 lookahead and M11 risk gating both consume this.
• Chain callout renderer + chain-vs-chain simulator scenario (#62, closes M8). Each chain template in Data/Strategies.lua now carries a labelKey = "CHAIN_<id>" field. 12 new locale keys (CHAIN_RMP_SAP_INTO_KIDNEY, CHAIN_RMP_FEAR_INTO_BURST, CHAIN_WMS_SHEEP_INTO_TRAIN, CHAIN_WLD_FEAR_INTO_CYCLONE, CHAIN_WLP_FEAR_INTO_HOJ, CHAIN_JUNGLE_TRAP_INTO_CYCLONE, CHAIN_BEAST_TRAP_INTO_INTERCEPT, CHAIN_TSG_HOJ_INTO_INTERCEPT, CHAIN_TRIPLE_CASTER_OVERLAP, CHAIN_RP_KIDNEY_INTO_BLIND, CHAIN_RD_KIDNEY_INTO_CYCLONE, CHAIN_SHATTER_NOVA_INTO_SHEEP) plus CHAIN_PICKED_PREFIX and CHAIN_STEP_PREFIX for the renderer, populated in both enUS and zhCN (86 keys per locale, parity gate green). StrategyEngine.Evaluate propagates labelKey and steps onto rec.chain and exposes the resolved link array via rec.chain.links. UI:Apply renders a chain block under the existing reason/callout subText: a localized title line with percentage + step count, then one indented line per step (GetSpellInfo(spellID) → spell name in-client, falls back to the category token in headless tests). UI prints the localized chain title to chat once per chain-id change — the placeholder per-step audio cue until M4 voice ships. New simulator scenario chain-vs-chain registered in Data/SimScenarios.lua (RMP enemy with DR-bumping cast sequence). Six new test cases verify catalog labelKey invariants, engine plumbing, UI narrate-once-on-change behaviour, and the simulator scenario runs without error.
• Chain scoring + engine integration (#61). Chain:ScoreAll(chains, opts) ranks a list of already-instantiated chains by ExpectedProb descending; opts.topK clips the output. Strategies:InstantiateChains(comp, primaryGUID, secondaryGUID, enemies) resolves a comp's chains templates into concrete chains by mapping byClass to an alive enemy of that class and targetRole ("primary" / "off-healer" / "off-healer-2" / "any") to a target GUID, dropping links whose caster class isn't on the field. StrategyEngine.Evaluate now picks the top-scoring chain and emits rec.chain = { id, label, expectedProb }. Configurable via db.strategy.chainK (default 3, parameter for future M10 opponent-response branching). WeakAuraBridge.GetChain() / GetChainId() / GetChainExpectedProb() exposed. The engine's reason text is unchanged — chain selection is a separate field so callouts in #62 can render it on its own UI line. Tests cover ScoreAll ordering + clipping, InstantiateChains byClass/targetRole resolution + link dropping + empty-result omission, and engine integration (rec.chain present + DR pre-bump degrades chain prob).
• Built-in chains per named comp (#60). Eleven comps in Data/Strategies.lua now carry a chains = { ... } field describing their canonical CC kill chains: RMP (sap-into-kidney, scream-into-burst), WMS (sheep-into-train), WLD (fear-into-cyclone), WLP (fear-into-HoJ), HUNTER_COMP (trap-into-cyclone), BEAST_CLEAVE (trap-into-intercept with scatter), TSG (HoJ-into-intercept), TRIPLE_CASTER (stacked fear+sheep), RP_2V2 (kidney-into-blind), RD_2V2 (kidney-into-cyclone), SHATTER_2V2 (nova-into-sheep). Each link is { spellID, category, byClass, targetRole } where byClass references a class in the comp's core and targetRole is a string ID ("primary" / "off-healer" / "any") resolved against the live state by future M8 wiring. 5 new tests in Strategies_spec assert the catalog invariants (>=10 chains, every link has a spell ID, byClass is in core, every chain validates against a fresh Chain state, every chain is well-formed). Hooks for #61 chain scoring + #62 callout renderer.
• CC chain primitive (#59, opens M8). New Chain.lua module. A chain is an ordered list of CC links { spellID?, target, category, by, castTimeS? }. Chain:Build(links) constructs one; Chain:Validate(chain) returns (ok, reason) after walking links and rejecting on DR-immune (reason="DR_immune"), pending caster CD (reason="cd_pending"), or empty input (reason="empty"). Chain:ExpectedProb(chain) returns the product of effective DR multipliers across links (0 if any CD is pending or DR has already hit immune). Within-chain DR accumulation is tracked so a chain of three STUNs on the same target correctly returns expected probability 1.0 * 0.5 * 0.25 = 0.125. Pure module; reads observation state from DRTracker / CooldownTracker without touching any WoW API directly. Foundation for #60 (per-comp built-in chains), #61 (chain scoring + lookahead), and #62 (CALL_CHAIN callout renderer).

Added

• Multi-reason burst gate (#73). StrategyEngine:BurstDecision(state, target, chain) returns { allowed, blockedBy, gates } with four named gates: kill_prob (passes when KillProb >= SE.BURST_KILL_PROB_THRESHOLD[aggression] — greedy 0.35, balanced 0.45, safe 0.55), chain_ready (passes when a chain with expectedProb > 0 is in play), incoming_pressure (passes when no DEFEND-level pressure: not healerUnderPressure, not enemyBloodlustActive, not multipleBurstsDetected), rating_aware (audit trail of the aggression label + numeric rating that influenced thresholds — always passes). blockedBy names the first failing gate in {kill_prob, chain_ready, incoming_pressure} order. Engine populates rec.burstDecision on KILL recommendations. WeakAuraBridge adds API.GetBurstDecision(). 7 new tests cover gate enumeration, kill_prob blocking on high-HP targets, threshold scaling with aggression, incoming_pressure blocking on under-pressure healer, all-gates-pass case, Evaluate population on KILL, and the WAB getter.

Added

• Kill-probability model with auditable breakdown (#72). StrategyEngine:KillProb(target, state) returns { prob, components }. Components: hp (1 − hp/100), defensiveDown (+0.10 when target's trinket has been used), immunityAbsent (+0.10 when no Ice Block / Divine Shield / BoP active), burstReady (+0.05 when our HoJ is up), healerLowMana (+0.10 when their healer is < 30% mana), drClean (+0.05 when target's STUN DR is fresh). Sum clamped to [0..1]. Weights exposed via SE.KILL_PROB_WEIGHTS for tuning. WeakAuraBridge adds API.GetKillProb(guid) and API.GetKillProbBreakdown(guid) so WeakAuras can render the probability per enemy. 6 new tests cover nil-target safety, monotonic-with-HP (100→50→10 strictly increases), component contributions (trinket-down, low-mana-healer, burst-ready), clamping to ≤1.0, and the two new WAB getters.

Added

• Rating-aware aggression (#71, opens M11). New db.strategy.ratingAggression config knob, default "auto". When "auto" and state.rating is known, derives aggression from bracket rating: <1800 → greedy, 1800–2200 → balanced, >2200 → safe. Explicit "greedy"/"balanced"/"safe" override; a number is treated as a rating override (handy for tests). Core:UpdateRating() queries GetPersonalRatedInfo() against the current bracket (returns nil headless). Core:CurrentAggression(state) resolves the active label; Core.PrepareAndEvaluate writes it to state.aggression before calling Evaluate. Three thresholds shift on the rating axis: SWAP score-gap threshold (greedy 0 → safe 20), defensive HP gate (greedy 30% → safe 50%), and the LOW_MANA_PUSH callout threshold (greedy 30 → safe 20). Recommendation gains rec.aggression and rec.rating for trace inspection. 9 new tests (4 in StrategyEngine_extra_spec, 5 in Core_spec) cover CurrentAggression resolution, UpdateRating no-API safety, defensive threshold flip on aggression, and low-vs-high swap behaviour at the same state.

Added

• UI polish bundle (#77). Compact mode: new db.frame.compactMode boolean default false. When toggled, UI:Apply hides the friendly + enemy icon rows so the recommendation block alone occupies the smallest possible footprint. Voice callouts: new Sounds.lua module maps callout keys to PlaySoundFile-compatible paths (Sound/Voice/<name>.ogg) and dispatches a one-shot cue per new top callout via the db.alerts.sound toggle. Headless-safe: when PlaySoundFile is unavailable, Play returns false without erroring. Audio assets ship as placeholder paths — the artist drop populates the binaries alongside the addon zip. 5 new tests (4 in Sounds_spec, 1 in UI_spec for compactMode).

Quality

• Confidence calibration audit (#76). New Tests/Calibration_spec.lua runs 100 deterministic synthetic states across HP / trinket / mana / DR axes, bins predictions into 10 deciles, and reports [CALIB] per-bin (predicted, ground-truth, error) lines for inspection. Headline metric: max per-bin error 0.10 for the v2.0 engine, well within the 0.20 budget. Engine adds SE:CalibrateConfidence(rawConf) — identity for v2.0; the hook is in place so future versions can ship measured bias corrections without touching callers.

Quality

• Benchmark suite (#75, opens M12). New Tests/Benchmark_spec.lua defines 21 canonical match scenarios spanning every major comp family (RMP / WLD / WMS / TSG / SHATTER / DRAIN / mirror / triple-caster / hunter-cleaves), plus mode-triggering edge cases (healer-trained DEFEND, enemy bloodlust DEFEND, multiple bursts, no enemies RESET, swap threshold, priest-dead-mid-match). Each scenario seeds a synthetic state and labels the expected mode (and optionally primaryTargetClass). Runner reports [BENCHMARK] agreement per scenario + overall rate to stdout for CI artifact capture. Soft floor of 50% so this is informational, not a hard CI gate. Current baseline: 81% (17 / 21) on the v2.0 engine.

Quality

• Rating-aware end-to-end test (#74, closes M11). New Tests/RatingAwareE2E_spec.lua runs the same synthetic state at rating 1400 (greedy) vs 2400 (safe). 4 cases: BurstDecision kill_prob threshold differs (low rating → lower threshold); defensive HP gate flips mode (greedy holds at 45 HP healer, safe goes DEFEND); ≥2 of {burst threshold, defensive HP, low-mana threshold} differ between low and high rating; blocked burst decisions cite the failing gate by name. Closes M11 — combined with #71 (rating-aware aggression), #72 (KillProb), #73 (BurstDecision gates), the engine now plays differently at different ratings, with every choice explainable via the audit trail. 524/524 green.

Quality

• Lookahead performance budget tests (#70, closes M10). Tests/Performance_spec.lua extended with a lookahead-on case asserting mean < 10ms / 99p < 30ms over 200 evaluations on a 5v5 / 3v3 state (target 3ms mean / 10ms 99p, with a 3x CI margin for noisy GH runners). Lookahead.lua gained a per-call response-distribution cache: EnumerateResponses is invoked once per Score call and reused across the K candidate chains (the distribution depends only on the profile). Lookahead:CacheStats() returns {hits, misses, total, rate} so consumers can inspect the hit rate; Lookahead:ResetCacheStats() zeroes the counters between measurement windows.

Added

• Pattern recognition for recurring kill setups (#69). New Patterns.lua module with five seeded patterns: RMP_CHEAP_BLIND, SHATTER_NOVA_SHEEP, FEAR_INTO_POLY, HUNTER_TRAP_SCATTER, HOJ_INTO_INTERCEPT. Each is an ordered list of { spellID, withinSeconds } steps. Patterns:Observe(spellID, ts) fed from Core.onCLEU on SPELL_CAST_SUCCESS; Patterns:Probability(id) returns completed-steps / total; Patterns:GetMatches(threshold) returns matched patterns (default threshold 0.7). Half-matches expire after STATE_TTL_SECONDS (12s default). Engine buildCallouts pushes CALL_PATTERN_<id> for every match. 5 new locale keys per locale (95 each, parity green). 12 tests cover catalog invariants, full sequence match, fractional probability, out-of-order rejection, window enforcement, TTL expiry, threshold gating, Clear, nil safety, each pattern's positive case, and unrelated-cast-in-the-middle resilience.
• /acc whatif counterfactual replay (#68). New slash subcommand: /acc whatif help, /acc whatif summary, /acc whatif skip <i>. Replays the current db.record.events through the engine with a single event removed (or other modifier) and prints how many recommendations diverged from baseline, with up to 5 sample diffs. Backed by two new exported helpers: Core:ReplayRecord(events, modifier) builds a synthetic state, runs the engine for each event, and returns a sequence of {mode, comp, chainId} snapshots — pure (snapshots + restores live CooldownTracker / DRTracker so it does not pollute in-game state); Core:DiffReplays(a, b) returns (count, samples). Two new locale keys HELP_WHATIF in enUS + zhCN (90 keys each, parity gate green). 6 new tests cover ReplayRecord non-leakage, DiffReplays identity / divergence with samples, /acc whatif with no record, help output, and skip output.
• Lookahead expectimax (#67, opens M10). New Lookahead.lua module. Re-ranks the top-K chains from M8 #61 by expected value over opponent responses (read from the M9 OpponentProfile when present, otherwise a 50/50 split). Bounded branching: defaults top-3 actions × top-3 responses × 3 plies = 81 leaves max per evaluation, configurable via db.strategy.lookaheadTopActions / lookaheadTopResponses / lookaheadEnabled. Engine integration: rec.chain gains an expectedValue field that reflects the post-lookahead score (vs the existing expectedProb which is the raw chain prob); expectedValue <= expectedProb by construction. lookaheadEnabled = false reverts to the greedy chain pick — useful for debugging / benchmarking. 9 new tests (7 in Lookahead_spec, 2 in StrategyEngine_extra_spec) cover EV computation, top-K clipping, profile-driven re-weighting (high trinket-prob lowers EV; low trinket-prob raises it), response probabilities summing to 1, and the disable toggle.

Quality

• End-to-end opponent-modelling test (#66, closes M9). New Tests/OpponentModellingE2E_spec.lua drives the full M9 pipeline (Signature → Update → Get → Estimate → buildCallouts) against synthetic teams. 4 cases: (a) CALL_SAVE_TREMOR_HOJ appears after 20 trinket-fear openings; (b) trained team A fires the callout while untrained team B does not, even with identical comp; (c) profile sanitisation — renaming the players produces a different signature, two independent profiles in db.profiles, but the same callout converges (the comp + observed behaviour determines the recommendation, not the names; no raw name strings ever appear in the persistent shape); (d) match-by-match: the callout first appears at observation ≥ 5 (the MIN_SAMPLES_FOR_OPINION gate). 471/471 green; locale parity green at 89 keys.
• End-to-end spec-match test suite (#58). Tests/SpecMatchE2E_spec.lua drives spec attribution through the real SpellSpecHints:Apply path (concrete spell IDs from Data/Spells.lua) and asserts Strategies:Identify picks the right spec-keyed variant. 15 cases covering each spec-keyed comp variant (RMP_DISC, SMR, WLD_RESTO, WLD_FERAL, SHATTERPLAY_SHADOW, SHATTER_FROST_2V2, HUNTER_PRIEST_BM_2V2), confidence calibration (1/3 vs 2/3 vs 1.0), wrong-spec disqualification (Holy priest → class-only RMP, not RMP_DISC), bracket isolation (a 2v2 shadow-priest setup never matches a 3v3 spec-keyed comp), and a catalog invariant that every spec-keyed comp's required spec is reachable from at least one SpellSpecHints entry.

Added

• Comp-match confidence scoring (#56). Strategies:Identify now returns (comp, confidence) where confidence is in [0..1]. Spec-keyed matches return 1.0 by construction; dynamic role-count comps (TRIPLE_DPS, DOUBLE_HEALER) return 1.0; class-only matches return the ratio of core-class enemies whose specGuess has been observed; legacy class-list-only callers (no enemies map) return 0.0. The recommendation gains two new fields — compConfidence and compSpecConfirmed — and the reason text appends a <COMP_ID> spec-confirmed|class-guessed (NN.NN) tag so /acc trace and /acc bugreport payloads show the confidence inline. UI subText renders a localized badge: <compLabel> (spec-confirmed|class-guessed). Two new locale keys — COMP_BADGE_SPEC_CONFIRMED and COMP_BADGE_CLASS_GUESSED — added in enUS and zhCN. WeakAuraBridge exposes GetCompConfidence() and GetCompSpecConfirmed(). Two-value return is backward-compatible — single-assignment callers silently drop the new value.
• Spec-keyed comp catalog (#55). Data/Strategies.lua comps gain an optional specs = { CLASS = "SPEC" } field. Strategies:Identify matches spec-keyed entries only when every required spec is explicitly observed via enemy.specGuess from spec inference; mismatched or unknown specs disqualify the spec-keyed entry so a class-only sibling declared later catches the fallback. Seven new spec-keyed variants seeded: SHATTER_FROST_2V2, HUNTER_PRIEST_BM_2V2, SMR_3V3 (Shadow Priest variant of RMP), RMP_DISC_3V3, WLD_FERAL_3V3 (no-healer variant, defaults to DEFEND), WLD_RESTO_3V3, SHATTERPLAY_SHADOW_3V3. Callers using the legacy class-list-only signature never match a spec-keyed comp because spec data only flows via the enemies map — so backward compatibility is preserved.
• Spec inference v2 (#57). Data/SpellSpecHints.lua expanded from 12 to 57 hints covering all 9 classes × 3 specs. Aura-applied hints (Shadowform, Vampiric Embrace, Spirit of Redemption, Moonkin Form, Tree of Life, Soul Link) and talent-implying casts (Vampiric Touch, Pain Suppression, Power Infusion, Holy Shield, Avenger's Shield, Repentance, Mana Tide, Tidal Force, Elemental Mastery, Shamanistic Rage, Shield Slam, Last Stand, Swiftmend, Mangle Bear, Siphon Life, Conflagrate, Shadowburn, Shadowfury, Arcane Power, Slow, Presence of Mind, Pyroblast, Combustion, Dragon's Breath, Icy Veins, Summon Water Elemental, Mutilate, Cold Blood, Blade Flurry, Adrenaline Rush, Premeditation, Shadowstep, Hemorrhage, Bestial Wrath, Intimidation, Silencing Shot, Readiness, Wyvern Sting) all carry definitive spec attribution. Core.onCLEU now routes SPELL_AURA_APPLIED / SPELL_AURA_REFRESH (not just SPELL_CAST_SUCCESS) through SpellSpecHints:Apply, so auras already in effect at the start of a match still teach the engine.
• WeakAura template pack (docs/weakaura-pack.md) — five copy-paste trigger templates (mode badge, burst gate, defensive alert, callout list, comp readout) that consume the _G.ArenaCoachTBC bridge. WeakAuraBridge.L(key) exposed so templates can resolve callout keys to the user's active locale without re-implementing the fallback chain.
• /acc record CLEU recording + tools/replay.lua. When record is on (default off), every CLEU event passed through Core.onCLEU is appended to ArenaCoachTBCDB.record.events (ring buffer, default cap 1000). tools/replay.lua <SavedVariables.lua> re-runs the captured log through the StrategyEngine offline and prints periodic recommendation snapshots, so a maintainer can second-guess specific calls without recreating the arena. /acc record on/off/status/dump/clear. enUS + zhCN HELP_RECORD strings.
• /acc bugreport error reporter. ErrorReporter.lua module with Capture(err, ctx), Recent(n), Reset(), Format(maxErrors), Sanitize(text), SetKnownNames({...}). Ring buffer caps at 20 captured errors. Sanitisation strips Player-XXX-XXX GUIDs, bare guid-... tokens, Name-Realm patterns, and any character names registered via SetKnownNames. /acc bugreport prints a markdown payload (addon version + client build + last 5 sanitised errors) ready to paste into a GitHub issue. enUS + zhCN HELP_BUGREPORT / BUGREPORT_HEADER strings.

Quality

• Locale parity gate added to CI (tools/check_locales.lua). Runs before the Lua test suite. Diffs every locale file's key set against enUS and exits non-zero with an explicit <locale> missing N key(s): listing. The existing Locales_spec.lua parity test stays in place as a second layer.
• Performance budget enforced as tests. Tests/Performance_spec.lua asserts StrategyEngine:Evaluate averages <5ms per call on a 5v5 state (target <1ms; the 5x CI margin tolerates noisy GH runners) and that 100 back-to-back simulated arenas stay within a 200kb GC delta (issue's 100kb target plus 2x slack for the spec framework). Catches scoring-loop regressions and enemy/cooldown table leaks before they ship.

Added

• Decision-trace logging. /acc trace on records every Evaluate recommendation (mode, target, reason, comp, bracket, callouts) into a ring buffer in SavedVariables. Defaults: disabled, cap 200 entries. /acc trace off, /acc trace status, /acc trace dump, /acc trace clear. Lets users (or me) post-mortem why the engine called a swap by inspecting the persistent log between sessions.
• Train detection. Core tracks damage events landing on our friendlies in a sliding window. When peelTriggerDamage (default 3) events arrive within peelTriggerWindow (default 5s), state.observations.healerUnderPressure = true and the engine forces DEFEND mode with reason trained. Both thresholds are configurable via db.strategy.peelTriggerWindow / peelTriggerDamage. Core._friendlyGUIDs is updated by every RefreshFriendlies so CLEU damage matching is fast.
• DR-aware callouts. buildCallouts consults DRTracker:NextMultiplier before emitting CC-related callouts. CALL_HOJ_KILL is suppressed when the kill target's STUN DR is in immune territory; CALL_CYCLONE_OFF is suppressed when the off-healer's CYCLONE DR is immune. No history = full multiplier = callout allowed (so the first cast still fires).
• Cooldown-aware scoring. SE.weights.kill_defensive_soon = -10 penalises kill targets whose major defensive (Ice Block / Divine Shield / BoP) comes off cooldown within ~15s. Catches the "we're about to waste burst into Ice Block" case. Uses observed casts from CooldownTracker; no history = no penalty.
• Mana-bar tracking on enemy healers. When an enemy healer drops below 25% mana, the engine adds the low_mana_healer = +20 weight to their score and emits the CALL_LOW_MANA_PUSH callout ("Healer low mana - push now" / "治疗蓝量低 - 压上"). enemy.manaPct was already populated by RefreshArenaEnemies; this wires it into scoring + callouts.
• M2 catalog data. 9 named 2v2 comps (RP / RD / Drainteam / Shatter / Enh+Priest / Hunter+Priest / War+Druid / War+Holy / SP+Pala) and 10 named 3v3 comps (RMP / WLD / Jungle / Shatterplay / LSD / RPH / Thunder cleave / Pala cleave / Ele Sham / Hunter+Lock+Priest) all tagged bracket = 2|3. Bracket-tagged comps win over agnostic ones when both match, so /acc enemy rogue priest in a 2v2 picks the RP_2V2 entry instead of falling back to the generic catalog.
• M2 bracket infrastructure. Core:UpdateBracket() reads GetBattlefieldStatus and sets state.bracket (2, 3, or 5). Hooked to UPDATE_BATTLEFIELD_STATUS, PLAYER_ENTERING_WORLD, and ARENA_OPPONENT_UPDATE so the bracket is fresh whenever the engine evaluates.
• Strategies:Identify(list, enemies, bracket) now accepts an optional bracket arg. Comps can declare bracket = 2|3|5 to opt into bracket-specific matching; bracket-tagged comps win over agnostic ones when both match.
• SE:GetWeights(bracket) returns the default scoring weights merged with any per-bracket overrides from SE.bracketWeights. 2v2 boosts role_healer to 40 (single-target healer kills win games); 3v3 raises it to 30. 5v5 uses defaults.
• WeakAuraBridge.GetBracket() exposes the current bracket to WeakAuras consumers.

Added

• M1 foundations: LICENSE (MIT), CONTRIBUTING guide, issue templates, PR template, manual smoke checklist.
• Release pipeline: .pkgmeta + .github/workflows/release.yml.
  • Every push to main auto-tags v{base}-dev.{run_number} and publishes a GitHub Pre-release with the addon zip attached and notes extracted from the [Unreleased] section of CHANGELOG.md. Pick up the latest testable build from the Releases page.
  • Pushing a stable tag v1.2.3 publishes a full release with notes from the matching ## [1.2.3] CHANGELOG section. Stable releases also upload to CurseForge / Wago when CF_API_KEY / WAGO_API_TOKEN are configured.
• ## Interface-BCC: 20504 directive in ArenaCoachTBC.toc so the packager builds a BCC-flavoured zip without a duplicate TOC.
• CooldownTracker now tracks Will of the Forsaken (7744, Undead racial) as a separate 120s cooldown. Surfaces via the existing CT:IsReady(guid, 7744) / CT:GetRemaining(guid, 7744) API. Aura-applied events are also caught, so WotF use is recorded even if the cast event is missed.
• Spells.CC_BREAK_RACIALS exposes the set of racials that act as fear/CC-breaks (currently just WotF). Engine consumers can iterate this when reasoning about "any CC-break ready" without hardcoding race-specific IDs.
• /acc selftest [verbose] — runs ~10 fast in-client assertions covering Spells data, CooldownTracker round-trip, DRTracker, StrategyEngine, Strategies/OwnComps identification, locale resolution, EventBus emission, and the WeakAura bridge. Reports SelfTest: N passed, M failed. Useful when a client patch or another addon clobbers state.
• SelfTest.lua module with Register / Reset / Run(verbose, printer) / RegisterDefaults. Composable so future modules can register their own checks.
• Spec inference v1. Data/SpellSpecHints.lua maps spec-defining casts to {spec, role} (Mind Flay → Shadow priest, Holy Shock → Holy paladin, Earth Shield → Resto shaman, Mortal Strike → Arms warrior, Bloodthirst → Fury, Crusader Strike → Ret, Mangle → Feral, Stormstrike → Enhancement, Lifebloom → Resto druid, Unstable Affliction → Affliction, Shadowform → Shadow priest). On every SPELL_CAST_SUCCESS for an enemy GUID, Core calls SpellSpecHints:Apply(enemy, spellID) to overwrite the default class-based role with observed evidence.
• 6 new max-rank TBC spell IDs in Data/Spells.lua to support the hint table (SHADOWFORM, MIND_FLAY, HOLY_SHOCK, EARTH_SHIELD, BLOODTHIRST, MANGLE_CAT). All sourced as TBC 2.4.3.
• /acc simulate <key> — scripted scenario runner. Sets up a fake enemy team, schedules events on C_Timer.After, and drives the live UI through a full fight without an arena. Three baked scenarios: rmp (Rogue/Mage/Priest opener with sap → poly → kidney → CS → fear → trinket), tsg-mirror (melee cleave training the priest), drain (Affliction lock + Disc priest 2v2 pressure). /acc simulate with no args lists scenarios; /acc simulate stop cancels a run in progress.
• Simulator.lua + Data/SimScenarios.lua modules. The simulator falls back to synchronous dispatch when no C_Timer is available, which is what unit tests rely on.

Changed

• Renamed S.PVP_TRINKET (which was incorrectly set to 7744 / Will of the Forsaken) to S.WILL_OF_THE_FORSAKEN. The actual PvP trinket effect remains S.PVP_TRINKET_EFFECT = 42292. No callers were affected — the old symbol was unreferenced outside the data file.

Notes

• An Anniversary-flavour TOC is deferred until the interface version for that client is confirmed (tracked in #8).
• Engine scoring (trinket_down = +20) still keys off the 42292 aura only. Wiring a separate cc_break_down weight for WotF is left to a follow-up — see #9 follow-up note.
• Spec-inference hints are conservative: only spec-defining casts are listed. Spells castable by every spec of a class (e.g. Mind Blast — any priest, Frostbolt — any mage) are deliberately NOT mapped because they would mislabel an enemy. Strategies:Identify already consumes enemy.roleGuess, so the next call after a relevant cast naturally picks up the new role.

1.1.0 - 2026-05-24

Added

• Dynamic own-team capability inference (Data/OwnComps.lua) with 5 archetypes (BURST_CLEAVE, SUSTAINED_CLEAVE, CASTER_CLEAVE, DRAIN_TEAM, BALANCED).
• 13-entry enemy comp database (Data/Strategies.lua) with per-archetype ownVariants: RMP, WMS, WLD, WLS, WLP, HUNTER_COMP, BEAST_CLEAVE, TSG, RLS, MIRROR_MELEE, TRIPLE_CASTER, DOUBLE_HEALER, TRIPLE_DPS.
• 25-getter WeakAura bridge exposed via _G.ArenaCoachTBC.
• CI workflow that runs all tests, computes coverage with luacov, and fails the build below 99%.

Changed

• Renamed addon from ArenaCleaveCoachTBC to ArenaCoachTBC. The name no longer implies cleave is the only archetype supported.
• Engine reads capabilities, not class names — adding a new class will not regress existing strategies.

Test

• 193 tests, 99.46% line coverage.

1.0.0 - 2026-05-23

Added

• Initial check-in: Core, StrategyEngine, CooldownTracker, DRTracker, EventBus, UI, Options, WeakAuraBridge.
• enUS + zhCN locales.