[2.1.1]

Added

• One universal build for every supported client family, selected by per-flavor TOC dispatch: Classic Era / Season of Discovery (11509), TBC Anniversary (20506), WotLK Titanforge (38002), WoW Forever (16001) and Retail (120100, 120105).
• Paladin seal twisting: twist window band, attempt/success feedback, lag calibration and an impossible-twist warning, driven by the live combat log on the classic client line.
• Paladin next-action guide: an on-screen prompt (Crusader Strike / Command / Blood / twist) with HOLD/RIDE overlay, rotation text and cooldown swipes; draggable and lockable.
• Twist Practice trainer: a self-contained simulation of seal twisting with its own control panel, /st practice, and TWIST!/MISS result feedback.
• Shaman spell-weaving assist: breakpoint marker, cast spark and clip warning for Lightning Bolt, Chain Lightning, Healing Wave, Lesser Healing Wave and Chain Heal, including Maelstrom Weapon stacks and live haste.
• WoW Forever support: the addon now drives its swing timers from the client's native PLAYER_SWING events on that client (no classic combat log required). Twisting attempt/success scoring stays classic-line-only (no combat log on Mainline-code clients): on WoW Forever the twist band and next-action guide run normally while the hit-rate badge shows a "twist scoring: classic-line only" notice instead of a 0/0 score (UNCONFIRMED in-client).
• Warrior, Druid and Hunter on-next-attack awareness: Heroic Strike / Cleave, Maul and Raptor Strike queue colours plus insufficient-rage indication where applicable.
• Hunter queued Steady Shot detection (opt-in, disabled by default).
• Sourced durations from LibClassicDurations where the library exists (Classic Era and TBC Anniversary), with hardcoded fallbacks everywhere else.
• A LibDataBroker-1.1 data source publishing the current swing timer for broker displays.
• An offline verification harness (bash tests/run.sh): syntax, per-client profile smoke tests, per-module state suites, TOC integrity, static wiring parity and luacheck.
• Mainline-client probes for spell knowledge (C_SpellBook.IsSpellKnown) and spell range (C_Spell.IsSpellInRange), consulted after the classic globals so rank resolution and out-of-range tinting work on Retail and WoW Forever; range keeps the 1/0/nil contract.
• Melee-weave proximity bar (weave_range.lua, default OFF): vertical fill plus yards label driven by the existing range poll; uses exact per-yard UnitDistanceSquared readings where the client exposes them (Era, TBC Anniversary, Titanforge, Retail, WoW Forever) with band fallback elsewhere.
• WoW Forever path: swing-engine reader hardened (missing/throwing speed reads keep cycles on schedule); probe gating, 3-return shape, C_Spell rank fallback, UnitDistance secret fallback, LRC fallback and swing-gated guides all verified.
• Retail path: all five guide engines return the empty display shape on RETAIL (defense in depth over swing-gating); HUD/rotation/design/options work with static bars.
• /st diag reports the C_Secrets secret-value restriction predicates (cooldown, aura, power, stats, cast) when the client exposes them, and api=no when it does not.
• TBC Anniversary guide models: hunter Kill Command / Multi-Shot weave / Bestial Wrath / Rapid Fire prompts (all TBC-gated, steady timing untouched); warrior Bloodthirst / Whirlwind cooldown rules with HS rage floor and Execute reservation (cleave slot replaced); paladin Judge-fit and Crusader-debuff cues, Consecration single-target suppression, CS-delay default 1.7s to 0.9s; shaman Maelstrom-gated weave markers, Stormstrike delay-if-fishing-Windfury, Fire Nova / Earth Shock slots and a totem-twist timing model; druid TBC chain (Rip at 4+CP expiry-only, no Rake, 40/42 energy gates, long-fight Rip over Bite, powershift SHIFT hold). Other clients byte-identical.
• TBC Anniversary completion: 0.05s player-aura poll (UNIT_AURA is unreliable on TBC); judgement marker line, standalone seal icons with cooldown swipes, twist-success glow; cooldown HUD slots + rotation text for hunter/warrior/druid/shaman with options and defaults; practice mode timeline, rotation display, keybind capture, haste toggles, rotation selector, proc icons.
• Classic Era correctness: hunter Aimed Shot prompt + Multi weave with Steady Shot gated off (no such spell on Era); druid Mangle gated off with an Era Shred/Bite/powershift chain; warrior Era fury (BT/WW, HS floor 30 + 2.5s buffer, Execute spam); paladin Era seals (SoR twist position, no Crusader Strike, Hammer of Wrath filler); shaman Maelstrom markers/rules enabled by observed aura (SoD rune) instead of family, so Era-base stays silent. Other clients byte-identical.
• Per-class completion: hunter Aimed Shot HUD slot + default + toggle; shaman totem-twist timing model live on TBC (GetTotemInfo wrapper, commit stamp + air-slot poll); warrior Rampage badge (stacks + upkeep window) with default + toggle; paladin Hammer of Wrath filler icon (8th guide slot); druid bear chain assessed and pinned per client.
• Wrath (Titanforge) + SoD min/max: per-class branches (hunter MM/SV + Chimera; warrior fury/arms + SoD runes; paladin FCFS + Exodin; druid SR/TF/Berserk + SoD; shaman MW5-only + Feral/Fire Ele/Magma/Lava Lash), Retail engine gates, practice scoring (grades, session score, persistent bests), TBC micro-details + Forever adaptations per class (effective weapon speed, HS-cancel, Slam-safe band, twist calibration, energy-tick phase, WF ceiling; Hawk/Sniper, Bloodthrill/Victory Rush, Twist-of-Light, Shifting Power, Farseer haste).

Changed

• All timing now runs on a single aligned high-precision clock behind the compatibility layer; every cross-client API access goes through one compat module.
• Spell, aura, cast, cooldown and haste reads are normalised across the classic and Mainline-code client families.

Fixed

• Warrior: the rage-vs-queued-cost colouring never ran because its event was never registered; Blood Surge glow could churn and double-release its pooled frame.
• Casting within the first seconds after a reload could error on the spellcast, current-cast, GCD and equipment-change handlers before the bar frames existed.
• LibRangeCheck-2.0 could be dereferenced while its checkers were still nil, and the melee checker could fail to resolve at all; both paths are now guarded.
• Shaman: the cast spark could jump backwards when a cast outlasted the remaining swing.
• Stale queue colours after combat (warrior, druid, hunter, shaman) and a shaman cast that a wipe interrupted without an end event.
• Paladin: the next-action guide read player mana through a bare global that modern clients may strip, and two twist timestamps used a raw clock instead of the compat clock.
• Retail / WoW Forever: restricted ("secret") readings could raise Lua errors in the cooldown, power, health, aura, cast, haste, attack-speed and inventory paths; every compat wrapper (plus LibGlobalCooldown, the Forever swing engine and the class consumers) now degrades to the documented unknown/idle shape and the swing engine keeps its scheduled cycles, with identical behaviour for unrestricted reads and on classic clients.
• bars.lua: the latency and GCD-length refresh walked a hardcoded three-hand list, so post_init crashed on every login for the single-hand classes (paladin, druid); it now walks the hands that exist. A duplicate convert_color in bars.lua shadowed the canonical converter and silently flattened every 0-255 colour to black.
• Hunter: the Multi-Shot cast-grace matcher indexed its name cache before the client's spell data had answered, crashing on a cold cache; the lookup is now guarded.
• LibRangeCheck-2.0: the LEARNED_SPELL_IN_TAB / PLAYER_TALENT_UPDATE registrations are guarded with C_EventUtils.IsEventValid (registering an unknown event throws on modern clients and aborted the library's activate()). Upstream 2.0 is deprecated; the guard mirrors the maintained fork LibRangeCheck-3.0, and tests/secret_guard_state.lua pins it against re-vendoring.
• LibRangeCheck-2.0's internals are hardened for the Mainline clients the event guard re-enables: the classic spellbook globals (GetNumSpellTabs / GetSpellTabInfo) and the IsSpellInRange / CheckInteractDistance reads no longer throw when the client dropped them, and the range poll has a dead-man that disables range checking with one notice instead of erroring every interval (tests/lrc_guard_state.lua pins both).
• TBC Anniversary Hunter Swing Engine: LibClassicSwingTimerAPI expansion level detection now correctly detects TBC Anniversary (LE_EXPANSION_LEVEL_CURRENT == nil in Lua UI space; GetClassicExpansionLevel() <= LE_EXPANSION_BURNING_CRUSADE / WOW_PROJECT_ID == 5), properly initializing the ranged_swing and next_melee_spells tables.
• Auto Shot & Steady Shot Interaction: Added Steady Shot ranks (34120, 25258) to noreset_swing_spells in LibClassicSwingTimerAPI so Steady Shot casts never reset the ranged cycle. Steady Shot can safely land within the 0.5 s Auto Shot cast window (red zone) without clipping Auto Shot, provided it finishes before the auto shot releases.
• Combat Log Noise Reduction: Added source/dest GUID filtering to ST:COMBAT_LOG_EVENT_UNFILTERED and LibClassicSwingTimerAPI:COMBAT_LOG_EVENT_UNFILTERED to only process events involving the player or current target/enemy, filtering out >95% of extraneous raid/combat noise while safely preserving incoming PARRY haste on the player.
• Hunter Bar Defaults & Visibility: Set Hunter mainhand, offhand, and ranged bar presets to show_condition = "either" and require_in_range = false, preventing swing bars from hiding or freezing at 1px when out of range or when rangefinder checks lag.
• Clock Drift & Progress Skew: Aligned CLEU event timestamps with the monotonic CanonicalGetTime() clock domain (_cleuClockOffset) and clamped swing duration and progress calculations in bars.lua and classes/hunter.lua so ends_at strictly leads start.
• Code Quality & Nil Safety: Guarded all nil-check warnings in bars.lua (set_gcd_marker_positions, on_bar_active, on_bar_inactive, set_gcd_width, set_deadzone_width, set_bar_texts, set_bar_color) and main.lua (ST:show_bar), fixed offSpeed global variable leak in LibClassicSwingTimerAPI.lua, fixed ui.practice_class_order nil guard and ui.toggle_timeline
• Cross-Class Bar Audit & Ranged Presets: Audited bar enable flags, weapon probes, and visibility gating across Warrior, Paladin, Rogue, Shaman, and Druid. Aligned Rogue and Warrior ranged bar presets to require_in_range = false (matching Hunter and ST.bar_defaults), ensuring manually enabled ranged bars are never unexpectedly suppressed by range-checking failures or melee proximity.

Guidance, HUD and design mode

• Cooldown HUD (ported from the archived build): a slot row with cooldown swipes, GCD dims, desaturate-on-cooldown, mana/out-of-range tints, proc glow and a trinket on-use/proc/ICD state machine; a per-class slot registry lets every class plug in its own abilities.

• Standalone rotation label (four formats, alignment anchors, fonts/outline/colour, visibility keys, drag/lock) and sound-preview dropdowns in the twist-sound options.

• Design mode: /st design draws a snap grid and alignment guides over every element (bars, HUD, rotation label, next-action guide, practice panel) with drag, double-click reset, and temporary lock override that restores each element's lock on exit.

• Next-action guides for warrior, druid, shaman and hunter on the paladin guide's provider contract (each with a pure decision engine and an offline suite), and the rogue guide on the same contract. The guides first shipped disabled "until verified in client"; that staging is over — all six guides now ship enabled by default (nextaction_enabled = true in every class default, config_presets.lua), and each can still be turned off per class.

• Class visuals: shaman breakpoint markers top/bottom, offhand cast spark, Windfury ICD strip, Flame Shock debuff bar, Stormstrike/Shamanistic Rage badges and shield-charge row; hunter Auto-Shot cooldown sync, auto-repeat gating, movement safe-stop maths, hidden-window cast bar with latency end-slice and clip tint, 4-state range helper, Rapid Fire bar and aspect icon; druid energy-tick column, form label and Mangle/Rip/Rake debuff bars; warrior Sunder Armor (with stacks), Deep Wounds, Rend, Thunder Clap and Demolition Shout bars, Shield Block bar, Slam cast bar and the Overpower flash (classic combat log).

• Hardening: the last direct cross-client global reads moved behind compat wrappers, the LibRangeCheck checker fallback/guards were tightened, the practice panel retries its frame creation if the addon finishes initialising mid-combat, and the guide engine's per-tick closure allocation was hoisted.

• Retail / WoW Forever: the 120105 (Retail 12.1.5) interface is now advertised alongside 120100, so the addon loads as 12.1.x (120100, 120105) with no out-of-date flag on either Retail patch (UNCONFIRMED in-client).

Build-out round + hardening (2026-10-17, verified offline, UNCONFIRMED in-client)

• Bloodlust / Heroism lust window: st.api.scan_player_aura single-pass 1..40 HELPFUL scan over ids 2825 / 32182 (TBC+ families only; Era/UNKNOWN closed), consumed by the hunter Steady-spam cue (H2, 0.5 s throttle) and the paladin twist-window narrowing (P3, 0.6x scale, nil-safe passthrough).
• Per-tick budget audit (C4): opt-in /st diag tick counters (diag_tick_counters, default OFF, zero per-tick cost when off, applies at next login/reload).
• Unified swing-safe overlay service (C3) + snap-to-row template registration (C6) + generic pips widget (D3); design mode gains snap templates and mouse-anchored dragging (nil = existing grab-offset behaviour).
• Paladin: live twist hit-rate badge (combat-only, optional session persist), JotC uptime mini-bar (21183, 20 s max-rank melee refresh), Art of War + Exorcism overlap pulse, seal-icon texture latch, adaptive CD icon row (Avenging Wrath / Divine Shield / Hammer of Justice / BoP / Holy Wrath / Lay on Hands + racial slot; Holy Wrath ranks and the racial list UNCONFIRMED).
• Warrior: bank cue (nil-off), composite BT/WW rage gate with HS reserve, Rampage tier badge, Battle Shout badge, Overpower glow-lifecycle fix, four arms/SoD HUD slot toggles (Rend / Mortal Strike / Bladestorm / Raging Blow, shipped true = no visible change).
• Hunter: lust-driven Steady-spam cue, Feign Death display reset, Readiness Rapid Fire refresh, safe-to-move cast zone, Bestial Wrath / Frenzy pet badges, Aimed Shot re-point prompt.
• Shaman: Stormstrike nature-vuln bar (17364, 12 s), Windfury-fishing SS HOLD tint + final-fraction wash, MW5 Lava Burst choice (SoD rune-gated) + Flame Shock collision guard, Rockbiter badge, shield-charge 3-charge standard fallback; round 2 adds 2-3-stack weave markers with safe-start zone band, per-spell clip state, Flurry ICD (0.5 s) + DW sync badges.
• Druid: Tiger's Fury / Berserk HUD slots (glow when ready), combo-point pips, EMA time-to-die, energy-cost countdown, GCD tick overlay, Omen glow, Maul badge, name-based Tiger's Fury resolution.
• Practice: generic scenario harness (C1) + P5/H5/W6/D7/S7 drills, panel reset button (re-anchors immediately), timeline drag persistence + close button, action-input paths for all four scenario shells.
• Enemy swing bar (opt-in, default OFF): hostile-target mainhand swing bar on the classic line only (Era / TBC / Wrath / Cata / Mists); inert on Forever / Retail / UNKNOWN.
• Options: 27 round toggles + 4 warrior slot toggles wired, all nil-safe (nil = OFF except hide-on-false HUD slots shipped true); hunter_multi_target ships with its toggle only — the multi-target engine is a future round; the shipped hunter_multi_target copy reads "Multi-target mode (manual, not wired yet)".
• Enterprise-grade wave: options-file warnings wired, practice panel/timeline/ scoring fixes, HUD provider-hidden slots now hide (no stale live-looking icons) with hidden-slot row reflow on hidden-flip outside combat (deferred only under lockdown; closed by the hud-reflow lane, bench/buildout_hud-reflow_2026-10-17.md), practice rotation + timeline design targets (registry 7 -> 9), rest-of-tree warning wiring per the trace-and-wire rule.
• Hardening (oracle R1-R4): Anniversary string-payload fallbacks on the paladin twist-success and warrior special-ability paths (numeric path byte-identical; owner /dump of the CLEU spellId slot type still settles it); secret-hole scan truncation closed in compat + paladin + warrior + shaman + druid aura loops (full 1..40 walks, nil = no match at this index); lust gate standardised on st.lust_supported == true (closed-on-unknown).
• Harness: 35 steps, 0 failed; luacheck 2 warnings / 0 errors (both KEEP-ruled standing). Every behavior above is verified offline only — the in-client checklist in TESTING.md is the owner acceptance pass.

Rogue sixth class (2026-10-19, verified offline, UNCONFIRMED in-client)

• Skeleton + wiring: NEW classes/rogue.lua (module open, empty id tables, HUD_SLOTS + get_hud_state, guide provider returning inputs_unknown=true, on_bar_update driving guide + HUD, on_combat_exit preserving the mainhand record) appended to all 7 TOC load orders after classes\warrior.lua; ST.ROGUE.defaults class keys (all opt-in OFF); ST.class_opts_funcs.ROGUE options section; tests/rogue_state.lua suite.
• Pure engine: NEW root rogue_guide.lua (pool-vs-spend on data-driven 65/85 pooling bands, Rupture-vs-Eviscerate decider on EMA time-to-die, per-client Envenom rule tables, EA-duty refresh, Shiv upkeep, Thistle-Tea deficit cue, inputs_unknown contract) with tests/rogue_guide_state.lua (206 checks).
• Cues + provider: proc-aware 2 s energy-tick predictor (Combat-Potency 15-energy grants never re-anchor; ambiguous tick hides, never guesses), energy-tick spark
  • pooling tint, Era-only post-mainhand-swing 0.1–0.2 s Sinister Strike window (sword mainhand only; never on TBC+), weapon-swap band ("swap: GCD + swing reset"), mainhand/offhand poison reminder (presence only — charges UNCONFIRMED, unrendered), live HUD slots (SnD/Rupture/EA/AR/BF/Tea/Kick) + POOL/NOW guide mapping with aura/energy/CP-unknown "…" rule.
• Practice drills: rogue_energy_pool (cast-start fraction vs tick marker, S..F bands, clip is F, nil tick unscored) and rogue_ss_window (Era 0.1–0.2 s post-swing offsets, ideal 0.1 s, F outside) scenarios; options copy states the unknown-hides rules for energy/tick, cues and the guide.
• Compat probes: api.GetItemInfo / api.GetItemCount / api.GetItemCooldown / api.GetWeaponEnchantInfo / api.UnitGUID moved behind probe chains in compat.lua (C_Item-first, classic fallbacks, secret/throw-safe unknown shapes); zero rawget probes remain in classes/rogue.lua. Latent fix: the offhand-enchant presence read is the 5th GetWeaponEnchantInfo return (four values per weapon slot), not the 4th — the 4th is the mainhand enchant id, so the old read could never show the poison reminder on a real client (bench/buildout_rogue-compat_2026-10-19.md).
• Options panel parity: SS-window / weapon-swap / poison-reminder toggles + colours, rogue_ea_duty opt-in, six guide thresholds pinned to the pure engine defaults, HUD slots, POOL/NOW copy (bench/buildout_rogue-options_2026-10-19.md). The ten Skeleton-lane cue keys (show_rogue_snd_cue et al.) were REMOVED by the rogue-parity lane (defaults + widgets + header/desc gone; absence pinned in tests/rogue_state.lua §§2/3 and tests/options_load_state.lua §9): the HUD slots + guide actions already cover every upkeep state the duplicate bar presentation promised, so it would have added no signal (desks/rogue-parity/). Harness floor held throughout: 37 steps, 0 failed; luacheck 2 warnings / 0 errors.
• Post-section landings: ST.ROGUE.get_rotation_key rotation-label provider (hunter shape; desks/rogue-parity/); ROGUE in diag.lua PROVIDER_CLASSES with guide/HUD pins (desks/rogue-diag/); Era Sinister Strike window ships RED and visible by default per owner order (show_rogue_ss_window true, rogue_ss_color red; desks/rogue-redzone/); warrior unmatched-payload counter with notice threshold (WARRIOR_PAYLOAD_NOTICE_AT; desks/five-class-gaps/); rogue-erroraudit hardenings F1/F2/F4/F5/F6 (NaN-safe number funnel, GetItemCount non-number check, POOL_BANDS / rule-order / rules-4-5-unknown-family pins; F3 accepted as documented residual — bench/rogue_erroraudit_2026-10-19.md, desks/rogue-auditfixes/).
• Rotation audits vs sims for all six classes (bench/*_rotation_audit_2026-10-19.md), followed by rotation-fix + follow-up waves: per-client EA/Tea/Shiv/builder rules (rogue), ICD discipline + SoD 200ms window (paladin), ww_ids→1680 + TBC HS ranks + hamstring engine (warrior), rule-7 opt-in + SoD slots + shock-capped ES (shaman), LnL arms + Serpent guard (hunter), trick/pooling + RipCP override (druid). Harness floor held throughout: 38 steps, 0 failed; luacheck 2/0.
• Operator-ruled cuts: Victory Rush SKIP (inert-by-default stands), hunter lust branch CUT (sim-faithful; alt-later opt-in documented), fight-remaining CUT, hawk-count CUT (probe-gated). W2 Battle Shout built fully; rogue H6 wait-HOLD + SoD ENVENOM rule wired; Blood Frenzy rune 412507 confirmed in-game by operator.
• Options-complete: per-class HUD template select (18 templates + Custom fallback), wowsims synergy copy in all 6 panels, LnL badge + Cold Blood + mana-threshold + RipCP widgets with live consumers; hard-wait WAIT render; shaman twist-mana feed; rogue 1CP/5CP Rupture DATA; druid FF fish rules + Claw slot; SoD paladin variants (seal-stacking, Shockadin, SoD-wrath) per fetched APLs; /st diag verbose gains id-resolution (67 wanted ids) + in-memory error log. All offline-verified, UNCONFIRMED in-client; see bench/live_test_script_2026-10-19.md.

Final hardening (2026-10-19, verified offline, UNCONFIRMED in-client)

• TBC-name / Era-id spell resolution across all classes with both-directions pins; compat hardened (C_SpellBook probes, CLEU mainline gate, shapeshift fallback); missing-class stubs for Monk/Demon Hunter/Evoker on Retail.
• Hunter melee-weave prog (distance band + LibRangeCheck drive, opt-in toggle) and trap tracking with Clever Traps handling.
• Usable-only trinket HUD (passive/unknown hide); running-GCD swipe on rogue; Claw HUD slot on druid; paladin guide→HUD slot mirror; warrior Sunder/Rend/stance-dance/hamstring rules.
• Guide/HUD/rotation-label/practice drivers contained under one error-quarantine discipline; /st diag verbose carries wanted-id resolution + in-memory error log.
• Settings pass: per-class copy/order/alignment, option for every shipped feature; HUD template gallery per class from researched WeakAura layouts.
• Release readiness audited (bench/release_audit_2026-10-19.md): versions 2.1.1, TOC coverage, harness 38/0, luacheck 2/0.

Practice/designer wave (2026-10-19; closes proctor G5)

• Top-level Practice page (ST.opts_table.args.practice, order 1.46): Start/Stop toggle on the same st.practice_ui.toggle session as /st practice and the class-panel buttons; Simulation rows (weapon-speed override, AW-macro pot/SoB, DST, queue window, BL delay) with live practice_ui consumers; rotation/timeline status text; feedback mirrors; desc enumerates all six drills (verified offline, UNCONFIRMED in-client; bench/paladin_practice_designer_2026-10-19.md, tests/paladin_state.lua page pins).
• Paladin drill deltas: twist sim + AW-combo drill (paladin_twist, paladin_aw_combo) on the shared session with start seeding (dst/bl/queue, nil = current defaults) and the archive-rule AW macro (pot sets the buff directly, SoB casts the full seal when the GCD is free) (verified offline, UNCONFIRMED in-client; tests/paladin_state.lua 520/520, tests/practice_state.lua registry pins).
• Warrior Execute drill: warrior_execute (20% HP threshold, BANK entry-rage bands S..F) grades live Execute-range entries with no sim of its own, plus the post-fight WarriorExecuteReport frame (verified offline, UNCONFIRMED in-client; tests/warrior_state.lua W6 pins, tests/practice_state.lua execute_phase pins; H6 ruling: report/live-log design stands, copy describes it honestly).
• Hunter forced-move drill: hunter_forced_move (cast inside a window is F, overrun into the next window is C, margin bands B/A/S; movement itself stays in-client) with panel shot icons casting simulated Steady/Multi/Aimed plus the shot report frame (verified offline, UNCONFIRMED in-client; bench/hunter_practice_designer_2026-10-19.md, tests/hunter_state.lua 969 §§P-U).
• Rogue drills: rogue_energy_pool (spend before the 2 s tick marker, clip is F, nil marker unscored) and rogue_ss_window (Era-only 0.1-0.2 s post-swing offsets, ideal 0.1 s, F outside; TBC+ runs pool only) (verified offline, UNCONFIRMED in-client; bench/buildout_rogue-pd_2026-10-19.md, tests/rogue_state.lua 678 §15).
• Druid powershift drill: druid_powershift (shift inside the 20-energy rule, fixed 40-refund Furor-5/5 sim constant, S..F bands) with the DRUID options Practice block and shared-session bridge (verified offline, UNCONFIRMED in-client; tests/druid_state.lua 602 §§30-35).
• Shaman weave drill: shaman_weave on weave.lua marker/clip math (safe-start margin bands S..D, clip is F, nil without the model) with the SHAMAN 5.x Weave Practice block (verified offline, UNCONFIRMED in-client; bench/shaman_drill_designer_2026-10-19.md, tests/shaman_state.lua 590 §§33-36).
• Both-ways start for every drill (/st practice, class-panel toggle, top-level Practice toggle into one session; panel icons + keybinds cast simulated spells) with S..F grades into the shared running score and persistent bests (verified offline, UNCONFIRMED in-client; tests/practice_ui_state.lua scoring/bests pins, tests/practice_state.lua 8-scenario registry pins).
• Designer: seal_icons + success_text descriptors (archive ELEMENTS shape; drag detaches, double-click reset re-attaches) and practice_hunter_report + practice_warrior_report descriptors (available only while shown); registry 3 bars + 2 paladin + HUD/rottext/NAG/practice panel/rotation/timeline + 2 reports = 13 design targets — 14 on a druid, which registers the Combo Points pips as a fourteenth (verified offline, UNCONFIRMED in-client; tests/designmode_state.lua count/title/by_key pins; the druid's own pips design key in classes/druid.lua).
• Reconcile: 6 SODOPT presence flips, design count 9->11 pins, hunter H. -> ST.HUNTER qualification, druid drill routing (verified offline, UNCONFIRMED in-client; desks/reconcile-pd/, 38/0 + 7/7 mutation kills on /tmp/opencode copies).
• Hardening: drill tick/cast quarantine, GCD nil-restore, sync fail-closed routing, toggle/post_init guards, use_shaman_drill accessor, report lockdown guard, ROGUE_DEBUG removal, top-level desc rewrite; G1-G3/G6 behavior pins landed, G4 stays filed (verified offline, UNCONFIRMED in-client; desks/hardening-pd/, 38/0 + 7/7 mutation kills on /tmp/opencode copies).

Rotation scenarios + per-class drills (2026-10-19, verified offline, UNCONFIRMED in-client)

• A shared rotation scenario engine (practice.lua): a run of ten GCD-chain decisions (H.ROTATION_RUN_STEPS = 10) where each step asks the class what its rotation wants and grades the cast. The four outcome classes are correct / out_of_order / miss / unknown; a correct decision grades S, a wrong or missing one grades F, and an unknown decision records no attempt at all (never a wrong grade). The run band is S 0 mis-picks / A 1 / B 2 / C 3 / D 4-5 / F 6+ on the mis-pick count (a miss counts as a mis-pick) — [DESIGN] addon policy, not a sim number (bench/practice_scenario_engine_2026-10-19.md).
• A practice mode select on the Practice options page: class (the shipped per-class session, the default), scenario (the class's own rotation, end to end) and drills (the class's focused drill list, one button each). A class that ships no scenario or no drills refuses that session with a clear message — it never falls back to another class's drill (bench/practice_scenario_engine_2026-10-19.md §2.2 and §3).
• All six classes ship a rotation scenario, and each answers from its own guide engine (ST.<CLASS>.practice_scenario_step), so the scenario teaches what the guide teaches and the two can never disagree. The six scenario ids: paladin_rotation, hunter_rotation, shaman_rotation, rogue_rotation, druid_rotation, warrior_rotation (practice_ui.lua PRACTICE_DRILL_KEYS; bench/class_features_drills_2026-10-19.md §2).
• 25 per-class drills ship, three to six per class, each grading a different skill (baseline resource behaviour is not a drill): paladin paladin_fcfs_priority / paladin_prepull / paladin_judge_fit; warrior warrior_core_order / warrior_ww_delay / warrior_sunder_upkeep / warrior_execute_burn / warrior_hs_queue / warrior_prepull; hunter hunter_shot_order / hunter_weave_clip / hunter_kill_command / hunter_prepull; rogue rogue_finisher_order / rogue_cp_finish / rogue_snd_upkeep / rogue_opener; druid druid_finisher_order / druid_powershift_rework / druid_roar_upkeep / druid_bite_execute; shaman shaman_enh_priority / shaman_mw_weave / shaman_stormstrike_hold / shaman_totem_upkeep (bench/class_features_drills_2026-10-19.md §§2.1-2.6; the per-class class files; the warrior's two 2026-10-20 additions — bench/warrior_hs_queue_2026-10-20.md §2.3, bench/warrior_apl_utility_live_2026-10-20.md §5).
• The eight built-in drills stay registered for backward compatibility; three are superseded by a class drill that owns the same skill (druid_powershift → druid_powershift_rework, shaman_weave → shaman_mw_weave, warrior_execute → warrior_execute_burn) and their grading is delegated to the class drill, so the retired truth never grades a player attempt. The class session wrappers now start the class's own drill, and the shaman's namespace-alias stopgap is gone (bench/practice_scenario_engine_2026-10-19.md §5; bench/superseded_wrappers_2026-10-19.md).

Practice gamification (2026-10-19, verified offline, UNCONFIRMED in-client)

• XP, levels and renown. XP per attempt is the live grade-point value (F earns 0 — a miss costs nothing, it just does not earn); the level curve is floor(100 * n^1.35), the cap is level 50 per class and past it a renown counter increments. Progression is per class plus a global account total, and a drill's XP is capped per day at 3x its median attempt XP, so one drill cannot be farmed for infinite levels (bench/practice_gamification_design_2026-10-19.md §2; bench/buildout_practice_gamification_foundation_2026-10-19.md).
• Per-drill mastery with tiers. A drill's mastery is the recency-weighted accuracy over its last 100 attempts (half-life 30); below 20 attempts it reads Unranked (one lucky S never grants mastery). The tiers are Unranked -> Bronze (60% / 20 attempts) -> Silver (75% / 50) -> Gold (85% / 100) -> Platinum (95% / 200) (bench/practice_gamification_design_2026-10-19.md §2).
• Streaks. A session streak (consecutive non-F attempts), a lifetime best streak, and a daily streak of consecutive calendar days with a completed session — a missed day freezes the daily streak, it never resets it (the operator's no-punishment ruling) (bench/practice_gamification_design_2026-10-19.md §2 and §9).
• Achievements. A static, data-driven id table — first_s, perfect_ten, hundred_attempts, all_drills_gold, level_10, week_streak, ten_hours — evaluated on every session end; each unlock fires a notice once (bench/practice_gamification_design_2026-10-19.md §4; bench/buildout_practice_gamification_foundation_2026-10-19.md).
• An unlock registry with three kinds — achievement, cosmetic and level-gated harder drill variant — and one hard rule: an unlock never changes how an existing drill behaves, a variant is registered as its own scenario. The seven achievements are registered; the cosmetic and drill-variant catalogues are empty in this build (the registry is the mechanism, not a shipped catalogue) (bench/practice_gamification_design_2026-10-19.md §5; bench/buildout_practice_gamification_foundation_2026-10-19.md).
• Stats. Lifetime aggregates per class and global: runs, attempts/hits/misses and accuracy (lifetime, rolling-last-20, last session), time active (counted only while a session is live, never while the panel sits idle), best session score, best single grade, best streak, XP/level/progress toward the next level, a per-drill table, and the last 50 sessions as a bounded ring buffer (bench/practice_gamification_design_2026-10-19.md §3; bench/buildout_practice_gamification_foundation_2026-10-19.md).
• A Progress options page (order 1.47, next to Practice): level, an XP bar, the per-drill mastery table, the achievements list, the stats summary, the all-classes line, a button that opens the stats frame, and a two-step reset with a cancel (bench/buildout_practice_progress_ui_2026-10-19.md §2).
• A stats frame with the accuracy sparkline over the last 50 sessions, the per-drill table, the per-class table and the session history (newest first, wheel scrolled with a tracking thumb), plus a Scope button that cycles the classes that have data and the account scope (bench/buildout_practice_progress_ui_2026-10-19.md §1).
• A panel level badge: level ("Lv 3", or "Lv 50 +2" past the cap), the XP into the level ("80/100"), the level-progress bar and the session XP ("+80 XP"), drawn under the panel so it overlaps nothing. It ships ON — the operator's request for visible levels/XP is the gate for that default-visible change — and the Progress page carries the toggle that turns it off (bench/buildout_gamification_visibility_2026-10-19.md; bench/buildout_options_pass_2026-10-19.md §1 item 4, the HU-9 widget).
• Persistence and repair. Progress and stats live in SwedgeTimerDB.profile.practice with a schema version: a corrupt blob repairs itself on load instead of throwing, a blob a newer build wrote is left untouched, and saves are throttled (session end, level-up, and at most once every five minutes during a live session — never per attempt) (bench/buildout_practice_gamification_foundation_2026-10-19.md; bench/buildout_practice_gamification_hook_2026-10-19.md).
• Progress is local-only: this build ships no settings export/import surface, so a profile cannot be moved between machines (bench/practice_gamification_design_2026-10-19.md §6, corrected 2026-10-20).

Audit fixes (2026-10-19, verified offline, UNCONFIRMED in-client)

• The error-ring install (F1). The seterrorhandler chain was never installed, so /st diag printed "errors none" as a fact when nothing could have been observed. The chain is installed once at init and removed again on logout, and the two empty states are now distinguishable: "errors none (handler installed, nothing thrown)" vs "errors not observed (no error handler installed)" (bench/audit_fixes_2026-10-19.md §F1).
• The offhand handler (F2). PLAYER_TARGET_SET_ATTACKING was registered but unreachable: has_oh was initialised false and never set again, so the handler always returned at its guard — and the test suite defeated the guard by hand, which is what hid the defect. The flag is now derived from the same offhand probe that fills the bar's weapon state, and the handler gained a ST:has_hand("offhand") guard: a paladin or a druid has an offhand weapon but no offhand bar, and without that guard the AceDB ** metatable auto-creates a synthetic table and the colour read raises inside a live event handler (bench/audit_fixes_2026-10-19.md §F2).
• The report's printer (F3) and the HUD's counter (F5). The diag report's only output path was the bare global print; it now falls back to the addon's own chat printer and says so once, so a report that changed channel is never mistaken for one that never ran. The HUD's combat-log registration counter was write-only; the runtime diag line now reads it (bench/audit_fixes_2026-10-19.md §F3 and §F5).
• Two dead helpers removed (F7, F11). ST:needs_gcd() had no callers and its synthetic-table read was a trap (the GCD underlay is gated per hand in bars.lua); the seven assert(event == event) tautologies are gone, with the unused arguments named instead (bench/audit_fixes_2026-10-19.md §F7 and §F11).
• The TBC name-vs-id resolution (F4). The family spell resolver now answers resolved / absent / unresolved instead of collapsing "the client could not answer" into "this spell is not here". Only absent latches a negative (it is final); unresolved latches nothing and the next attempt asks the client again, behind a five-second backoff so a client that cannot answer never re-walks the rank table every frame. The druid and the warrior class latches moved onto the state API (bench/f4_family_spell_id_2026-10-19.md; bench/class_f4_latch_2026-10-19.md).
• The base TOC (F6). The base SwedgeTimer.toc no longer advertises a mainline interface: it is the classic-line fallback (11509, 20506, 38002), because the vendored LibClassicSwingTimerAPI registers a combat-log event at file scope and throws on mainline-code clients. The mainline flavour ships without that library (bench/audit_fixes_2026-10-19.md §"Still open"; the base toc's ## Interface: 11509, 20506, 38002).
• One shared error-isolation helper (F10). The isolation contract (count -> report once -> quarantine after 25 -> the caller's own degrade) was implemented six times with six quarantine constants, and the copies had already diverged (one printed on every refusal, the other five printed once). A new root module guard.lua now owns it once — the shared entry points (st.guard.run(owner, site, fn, ...) and its note / quarantined / refusal siblings) — and the six call sites (bars.lua, hud.lua, nextaction_ui.lua, practice_ui.lua twice, rotationtext.lua) were migrated onto it; st.guard.QUARANTINE_THROWS is the only threshold constant left (bench/f10_shared_guard_2026-10-19.md).
• The scenario-context crash. The default scenario context raised on an undefined SCENARIO_RESOURCE_KEYS constant for the druid, warrior and shaman; the list is now a local, verified against the six classes' own gather_guide_inputs, and the lift can no longer crash (bench/f10_shared_guard_2026-10-19.md).
• The class-module namespace fix (HU-7). The pure half resolved the class module from the namespace table, but the class tables live on the AceAddon table — so the engine found no class module at all. The AceAddon table is resolved first now and the namespace is the fallback, which is what let the shaman's st.SHAMAN = ST.SHAMAN stopgap be removed (bench/practice_scenario_engine_2026-10-19.md §6; bench/superseded_wrappers_2026-10-19.md §1.3).

Closure wave (2026-10-19, verified offline, UNCONFIRMED in-client)

• The Wolfshead probe (druid). api.is_item_equipped(8345) — a tri-state read (true worn / false readable-and-absent / nil unknown, never a guessed false) — is published as inputs.wolfshead_equipped and throttled to one read per second (it was ~380 equipped-slot reads/s for an input no rule reads yet). It is a reserved plug-in point: the Furor half of the powershift basis is still unproven, so no rule consumes it (bench/compat_probes_2026-10-19.md §1.2; bench/buildout_closure_triad_2026-10-19.md §3).
• The Serpent target scan (hunter). inputs.serpent_active is no longer a hard-coded nil: it is a real three-state feed from a target debuff scan (the name "Serpent Sting" first, the sourced rank ids second), so the Chimera is withheld only when the sting is known to be down, and an unreadable debuff still allows the prompt (bench/hunter_serpent_2026-10-19.md).
• Call of the Elements + the per-family Stormstrike (shaman). The water-totem clock (client-sourced slot 3, tri-state: the seconds left, 0 for an affirmatively empty slot, nil for an unknown read) feeds a new 18th guide slot — { key = "coe", spell_id = 66842 } — which fires on Wrath while the water totem has under 20 s left, plus a display-only HUD mirror that hides unless its toggle reads exactly true. The Stormstrike swipe fallback is now per family (Era 20 s / SoD 6 s / TBC 10 s / Wrath 8 s, 10 s elsewhere — no invented number) on both the guide slot and the HUD mirror (bench/shaman_coe_ss_2026-10-19.md).
• The Flame Shock default is sim-exact. guide_flameshock_refresh now ships 0 (every sim re-applies Flame Shock only when the dot is inactive) instead of 3, and the gather's stale or 3 fallback was aligned to the same value — the shipped 3 shaved about three seconds off every Flame Shock dot (bench/buildout_options_pass_2026-10-19.md §2 item 3; bench/buildout_closure_triad_2026-10-19.md §2).
• The Hemorrhage builder (rogue) + the swapped-id corpus fix. The builder candidate chain is now { "Saber Slash", "Mutilate", "Hemorrhage" } ahead of the dagger Backstab and the Sinister Strike fallback, with a family gate (Hemorrhage is an Era/TBC/Forever builder, never a SoD one — it never sits ahead of a rune the sim prefers) and unknown-hides (a name the client does not know falls through, never a wrong builder). The bench corpus had the TBC ids swapped: 26862 is Sinister Strike and 26864 is Hemorrhage (desks/rogue-hemorrhage/journal.md; bench/filed_closure_sweep_2026-10-19.md §9-N1).
• The Forever Echo and Holy Strike ids (paladin). The two deliberately-empty Forever id tables now carry 1311703 (Echo of Command) and 1311704 (Echo of Righteousness), and Holy Strike 10333 is published as a plug-in point. All three are web-confirmed, client-unconfirmed and name-anchored: an id only counts when the client's own resolved name matches, so a wrong id leaves the input unknown instead of producing a wrong prompt (bench/buildout_paladin_bs5_2026-10-19.md).

Scenario surface + APL utility layer (2026-10-20, verified offline, UNCONFIRMED in-client)

• The scenario surface works now. Two CRITICAL defects were offline-invisible and are fixed. The paladin's scenario mode graded nothing in-game: its own context builder was written for the twist simulation and served that shape to a live call, so every sim-shaped field was nil, the decision was nil and every step was UNKNOWN — no attempt, no grade, no XP. The builder serves both shapes now (the sim shape verbatim; the live shape from the class's own gathered guide inputs). The drill list answered UNKNOWN on every press for five of the six classes: the row built the context with no pick, and every pick-based grader answers nil without one (the hunter was the one class that worked — its grader reads the context's own pick). The drill-list row now carries the class's actions as the pick channel (bench/paladin_scenario_surface_2026-10-20.md §1; bench/practice_ui_fixes_2026-10-20.md §1).
• The shaman's and the druid's own sessions are reachable again. The supersession was checked before the shipped session, so the drill list won, the class panel's icons hit a nil session and were silent no-ops. The shipped session is checked first now (a class with a session of its own keeps it, and that session grades through the class's own drill), and any drill cast with no session is loud — it prints the reason and counts it (bench/practice_ui_fixes_2026-10-20.md §2).
• The scenario row no longer leaks frames. The row rebuilt its button set on every step and only hid the old buttons — roughly 200-300 hidden frames per run. The buttons are reused from the row's own pool now, so a rebuild is a pass, not a new set (bench/practice_ui_fixes_2026-10-20.md §3).
• The rotation scenario is gated on a live swing engine. The Retail client keeps the inert stand-in and cannot run a timed rotation, so the session is refused loudly — the refusal reads "Practice scenario is not available on this client (no live swing engine)." (practice_ui.lua) — and the Practice page's availability note carries its own, separate string: "The rotation scenario cannot run on this client (no live swing engine)." (option_tables.lua). Two strings for one behaviour. Only the explicit stand-in counts as dead — a client that has not reported its engine yet is unknown and is never blocked (bench/practice_ui_fixes_2026-10-20.md §4).
• The drill counters die with the session. They hung off the addon object, so a band's mis-picks survived into the next session and a drill decayed S → F for the life of the profile; the drill-list session owns its own identity and window now (bench/practice_ui_fixes_2026-10-20.md §5).
• The twist minigame is the paladin's. The twist drill leads the paladin's drill list (the flagship practice stays reachable from the paladin's panel), and a class that publishes no drills of its own and owns no twist drill is refused loudly ("No practice for this class.") instead of inheriting the paladin's minigame. The warrior is not one of them — it publishes six drills of its own (bench/practice_ui_fixes_2026-10-20.md §6; the operator's ruling; the warrior's own drill keys in tests/warrior_state.lua).
• A hold is a decision; unknown is not. The class contract grows an explicit hold statement. The engine grades a cast during a hold as an illegal cast (F — it counts as a mis-pick) and a correct hold records no attempt (no points, no streak, no mastery tick); a step with neither an action nor a hold stays UNKNOWN. The paladin (the twist RIDE window and the FCFS "nothing is ready"), the rogue (POOL), the shaman (RIDE / WAIT), the hunter (RIDE / HOLD) and the warrior (RIDE / HOLD) state their own holds. The druid's holds came in two shapes — nine spell-naming holds (a decision to cast) and one press-nothing hold (RIDE, the bear chain's queued Maul); the split is documented in the Audit-fix wave section below, which supersedes the earlier note that the druid stated no hold at all (bench/shared_engine_bs8_2026-10-20.md §1; bench/class_hold_statements_2026-10-20.md §2 — superseded by bench/druid_hold_ride_2026-10-20.md §7).
• The APL utility layer. Every class's rotation table now carries the sims' full priority list, not just the damage rows. Two rules apply cross-class: the rotation table carries the APL's spell rows while the APL's pre-pull rows live in a per-class pre-pull table (ST.WARRIOR.PRACTICE_PREPULLS, ST.HUNTER.PRACTICE_PREPULLS, ST.ROGUE.practice_prepulls, ST.DRUID.PRACTICE_PREPULLS, ST.SHAMAN.PRACTICE_PREPULLS, ST.PALADIN.practice_prepull_rows); and every row carries two flags — live (the row's read exists and the line can be evaluated) and engine (the frozen engine has an arm that can ask for it). A row with no read is never asked for; a row with no arm is never a button that can be the right answer (bench/rotation_gap_vs_apl_2026-10-20.md §0.0 and §9; the per-class build and live artifacts).
• Warrior. Battle Stance (2457), Berserker Stance (2458), Gladiator Stance (412513), Bloodrage (2687), Battle Shout (Era/Forever 25289, TBC 2048), Recklessness (1719), Death Wish (Era/Forever 12328, TBC 12292), Sweeping Strikes (12723), Berserker Rage (18499), Blood Fury (20572), plus the WotLK rows Thunder Clap (47502), Shattering Throw (64382) and Heroic Throw (57755). Twelve of the utility spine's thirteen rows carry an engine arm; Gladiator Stance is engine = false — the SoD APL gates it on the T1 damage set bonus, and the client can answer the Forecast buff but not the set ownership. (The compiled order's cleave row has no arm either — no rule ever outputs it) (bench/warrior_apl_utility_build_2026-10-20.md §1; bench/warrior_apl_utility_live_2026-10-20.md §§1, 4; bench/warrior_utility_residuals_2026-10-20.md §1).
• Hunter. Rapid Fire (3045), the Serpent Sting upkeep (25295 / 49001), the aspects (Viper 34074, Hawk 27044, Dragonhawk 61847, Falcon 469145), Kill Shot (the SoD rune 409593, WotLK 61006), Silencing Shot (34490), Explosive Trap (49067), Multi-Shot (49048), Black Arrow (63672), Raptor Strike (14266), Hunter's Mark (14325), Mongoose Bite (14271), Immolation Trap (14305), Wing Clip (14268), Arcane Shot (14287) and Wyvern Strike (458482) — one row per family, each at the APL's own priority position (bench/hunter_apl_utility_build_2026-10-20.md §3; bench/hunter_utility_live_2026-10-20.md §1).
• Rogue. Vanish (SoD 1856, WotLK 26889, Forever 1857), Preparation (14185), Premeditation (14183), Ambush (11269), Garrote (11290), Venom (1310703), the SoD Poisoned Knife (425012), Tricks of the Trade (57934), Killing Spree (51690), Blood Fury (20572), Elune's Light (1259799) and the Gnome racial Eureka! (1259812). The Cloak of Shadows row was refuted — 1857 is Vanish rank 2 on the TBC and Forever clients and the spell does not exist on either (it was removed in 2.3) (bench/rogue_apl_utility_build_2026-10-20.md §1 and §3; bench/rogue_apl_residual_reads_2026-10-20.md §§1-2).
• Druid. Cat Form (768), Leader of the Pack (17007), Faerie Fire (Feral) (17392 / 27011), Innervate (29166), Elune's Light (1259799) and the TBC Bloodlust row (2825), with Claw (9850) left RESERVED (the read that decides "the Shred cannot land" does not exist) and the Bloodlust row honestly live = false — 2825 is the Shaman's Bloodlust, a spell the druid cannot cast, and the APL's own "Bloodlust Time" variable has no addon counterpart (bench/druid_apl_utility_build_2026-10-20.md §3; bench/druid_apl_residuals_2026-10-20.md §§4-5).
• Shaman. Bloodlust (2825, WotLK), Strength of Earth (Era 10442 / TBC 25528 / Forever 25361), Mana Spring (25570, TBC) and Frost Shock (25464, TBC), each at the APL's own position; the Lightning Shield maintain now covers SoD and Forever as well, and the Frost Shock filler is a real, class-owned option (guide_filler_shock, default earth — the APL's own default) instead of a dead variable (bench/shaman_apl_utility_build_2026-10-20.md §1; bench/shaman_residuals_2026-10-20.md §§1.1-1.2).
• Paladin. Every rotation row now carries its APL citation, its family gate, its per-family id and its TBC name. The rotation table carries 17 rows: the twelve shipped ones plus the APL utility layer's five — Hand of Reckoning (67485) and Hammer of Wrath (48806) (the WotLK APL's top two rows, both with an engine arm), plus Judgement of Wisdom (53408), Holy Strike (10333) and Divine Favor (20216), which are live (their read exists) but carry no engine arm and so never appear as a button. The guide carries twelve slots (the two band rows are new) and the pre-pull table carries eleven rows — TBC (Sanctity Aura 20218, Seal of Command 20375, Seal of the Crusader 27158, Judgement 20271, Seal of Blood 31892), Forever (Retribution Aura 10301, Seal of Command 20920, Seal of the Crusader 20308) and SoD (Seal of Command 20920, Seal of Martyrdom 407798) — with no pre-pull row ever in the combat rotation. The spell names are the table's own labels (ST.PALADIN.practice_prepull_rows): 20920 is Seal of Command (the class's own seal ladder puts it there, and the Forever client's own ladder agrees) and 20308 is Seal of the Crusader. Divine Storm and Holy Shock are different spells in this tree — Divine Storm is 407778 (SoD) / 53385 (WotLK) and Holy Shock is 20930 — and neither is a row of this table. One label is contested inside the tree: the class's own sob rotation row names the same SoD 20920 r5 "Seal of Martyrdom" (that row's own SoD id is 407798), so the two spell names for that id disagree between the pre-pull table and a source comment — this entry follows the pre-pull table. The two WotLK band rows also ship as HUD mirrors: hud_show_hor and hud_show_how_wrath, both off by default and WotLK-only, and the paladin's three HUD templates now name all fourteen shipped slots (the band rides P-T2 "Full Ret + Trinkets" and stays off in P-T1/P-T3). A correction to the earlier release notes: the pre-pull table was recorded here as carrying two rows (Sanctity Aura on TBC, Retribution Aura 10301 on Forever) — it carries eleven (bench/paladin_apl_utility_build_2026-10-20.md §§2.1–2.5; bench/paladin_hud_utility_band_2026-10-20.md §§1.2–1.3; bench/paladin_audit_close_2026-10-20.md §§1.2–1.3).
• The warrior's Heroic Strike / Cleave queue drill (warrior_hs_queue) — the operator's mechanic, with the sim's own citation: the queue lifts the dual-wield miss penalty (DisableDWMissPenalty), and the penalty it lifts is +19 % over a 5 % white base. The drill grades the press window — the gap between the mainhand's last swing and the offhand's next swing — on the tighter of the two sides (S ≥ 250 ms, A ≥ 150 ms, B ≥ 80 ms, C ≥ 20 ms, D inside the window, F outside it). The sim's 250 ms is its own reaction-time model, not a client fact: the in-client window is UNCONFIRMED (bench/warrior_hs_queue_2026-10-20.md §§1-2). Correction to the earlier release notes in this section: the first documentation pass recorded this drill as absent from the build; it landed on 2026-10-20 and that note has been removed.
• The corpus was corrected — do not trust the old ids. The rotation-gap analysis is the seed of most of the wrong ids in the corpus, and the five utility lanes re-fetched every APL file and every sim spell table and disproved 31 ids, names and gates. The corrections are applied in place in bench/rotation_gap_vs_apl_2026-10-20.md (§0.0 lists each one with its old value, its corrected value and its evidence; §9 summarises them), and sixteen other bench files were corrected the same way. The headline ones: 17364 = Stormstrike (no wowsims shaman sim has a Nature's Wrath ability at all), 1857 = Vanish rank 2 (not Cloak of Shadows), 9850 = Claw rank 5 on Era, TBC and Forever — and the Forever client has no Mangle spell at all, 2825 is the Shaman's Bloodlust, the WotLK hunter's Kill Shot is 61006 and its Explosive Shot 60053 (53351 / 53301 are the client's rank-1 fallbacks), the TBC Steady Shot is 34120 (56641 does not exist on the TBC client), and the Forever client has no Steady Shot at all — its row was removed (bench/rotation_gap_vs_apl_2026-10-20.md §0.0 and §9).
• The Forever druid chain no longer names a spell the client does not have. The Mangle branch answered "mangle" on every tick on the Forever client (the spell does not exist there, so the Mangle timer is always nil), which made the Bite and the builder branches unreachable; the branch was removed and the chain now follows the APL's own order (bench/druid_apl_residuals_2026-10-20.md §1).
• Offline state at this landing: bash tests/run.sh → 41 steps, 0 failed; luacheck → 2 warnings / 0 errors in 83 files (the standing floor; both warnings are pre-existing, in files this wave never touched). Every behaviour above is verified offline only — the in-client checklist in TESTING.md and the operator script in bench/live_test_script_2026-10-19.md are the owner acceptance pass.

Audit-fix wave (2026-10-20, verified offline, UNCONFIRMED in-client)

The post-everything deep audit found two HIGH and four MEDIUM defects in the practice surface, the paladin's rotation slice and the compat contract. All six are closed. Every entry below is verified offline, UNCONFIRMED in-client — the operator's session is the only thing that can confirm the in-client behaviour.

• HIGH — the druid's press-nothing hold was graded as an action (the bear's queued Maul). The druid was the only class of six that mapped the guide's HOLD onto a synthetic press: every hold text became an action key, so the guide's RIDE (the bear chain's queued Maul — the guide already said no action) turned into a "ride" button. Both directions were wrong: pressing nothing (correct) read as a miss → F, and pressing the rendered "ride" button graded S with points — points for idling, which the engine's hold contract forbids. The class's own priority drill also banked a false mis-pick for every real pick. The guide's hold vocabulary is now split in two, as two published tables: ST.DRUID.PRACTICE_HOLD_KEYS (the nine spell-naming holds → the action to press) and ST.DRUID.PRACTICE_HOLD_PRESSES_NOTHING (the one press-nothing hold, RIDE). The scenario step states the press-nothing hold as { action = nil, hold = true, reason = "RIDE" } — a cast during it is an illegal cast (F), doing nothing during it is correct but worth no points — and the priority drill skips hold steps, so a hold banks no mis-pick (bench/druid_hold_ride_2026-10-20.md §§1–3).
• HIGH — a completed practice session was lost on logout. ui.end_progress_session was the only writer of the completion bonus, the daily streak (class and account), the run counter, the accumulated time active and the bounded history row — and it had exactly one caller: the session's stop. PLAYER_LOGOUT did not flush it, so a player who practised and then logged out kept the per-attempt XP and lost everything else, with no error and no notice. A dedicated ui.flush_progress_session(self, now) now ends and persists the progression session without tearing the frames down (a logout is not a session end), is idempotent, and reports honestly whether it wrote. It has three callers: the session's stop, main.lua's PLAYER_LOGOUT, and the practice panel's hide. The write lands in the saved variables because AceDB registers its own PLAYER_LOGOUT handler at library load — before this one — so it has already stripped the defaults when the addon writes the fresh blob into profile.practice (bench/practice_persist_2026-10-20.md §"H2"). The caveat is stated, not hidden: a UI /reload fires no PLAYER_LOGOUT and writes no SavedVariables — the client serialises them at logout only — so a mid-session reload loses the in-flight session whatever the addon does. That is client behaviour, not an addon defect (bench/practice_persist_2026-10-20.md §"H2", caveat).
• MEDIUM — a Progress-page reset mid-session dropped every later attempt. The reset nils the hook lane's live cache, and the attempt recorder then returned before the profile was re-created — so the score kept moving while no XP, mastery, streak, stats or badge did, with no error. Two halves: the attempt recorder now ensures the profile itself before it records, and the reset also clears the in-flight session's counters (the session keeps running; only the progression bookkeeping restarts, and the erased XP comes off the badge). The session's start stamp is deliberately not reset, so the run's time active still matches the score summary the history row carries (bench/practice_persist_2026-10-20.md §"M1").
• MEDIUM — the paladin's practice rotation slice ignored the family gate. The slice offered the TBC client a "Hammer of Wrath" button the TBC rotation can never ask for — the how row declares ERA/SOD/FOREVER, and the engine's execute arm is Era/Forever only — so pressing it graded out_of_order → F for a spell the rotation never wanted. The slice now applies the same families predicate the rogue, warrior and druid slices use, on both paths (the order's own keys and the APL-row path). The family token comes from the order, not the client: the SoD game runs on the Era client, so the SoD variants' own orders name SOD. The per-client slices after the gate: TBC drops how; Era drops cons/exo from the rotation part of the strip (the corpus's family gate — the Era APL carries no filler row) while the presses stay reachable through the panel's asked-for vocabulary; Forever drops cs/sob; SoD and WotLK are unchanged (bench/paladin_audit_close_2026-10-20.md §1.1).
• MEDIUM — the spell-knowledge contract did not match the code. compat.lua's comment claimed a client with none of the probes reports "unknown" instead of erroring; the code answered false. Every current caller treats false as fail-closed, so the behaviour was safe, but the contract was wrong: a caller could not tell "no probe answered on this client" from "the player does not know the spell", and a row that must hide when the client cannot answer had no way through the API. The boolean stays binary, never nil (a nil answer was proven not behaviour-neutral for two raw consumers), the comment now states the contract the code implements, and an additive api.is_spell_known_state(spell_id) answers the tri-state true,"known" / false,"not_known" / false,"unknown". No existing caller was touched, and the paladin's WotLK HUD band gate now reads that state, so "the player does not know it" is distinguishable from "this client cannot answer" (both still hide — unknown hides). api.is_spell_known itself is unchanged (bench/compat_m3_contract_2026-10-20.md §§1–3; bench/paladin_audit_close_2026-10-20.md §1.5).
• MEDIUM — a client without a calendar froze the daily XP cap forever. current_day() returned nil when the client has no date, and the pure layer resets the drill's daily cap only when the day ordinal is non-nil and the day changed — so the counter accumulated across days until the drill awarded 0 XP for good, silently. Two layers: the day falls back to the client's own clock (floor(time()/86400) — an opaque day counter whose consecutive days differ by exactly 1), so the cap still resets every day and the daily streak still advances; and a client with neither a calendar nor a clock now prints one notice naming the cause instead of silently awarding nothing. The notice goes through the addon's own printer, not the badge-gated announcer, because a diagnostic that rides an opt-in toggle is a diagnostic nobody sees (bench/practice_persist_2026-10-20.md §"M4").
• Dated history note (2026-10-20). The "Scenario surface + APL utility layer" entry above originally recorded the druid as stating no hold at all; that wording was corrected in place when the HIGH defect above was fixed, so no correction is owed to it now. The druid's guide holds in two shapes — nine spell-naming holds (a decision to cast) and one press-nothing hold (RIDE, the bear chain's queued Maul) — split by ST.DRUID.PRACTICE_HOLD_KEYS and ST.DRUID.PRACTICE_HOLD_PRESSES_NOTHING (bench/druid_hold_ride_2026-10-20.md §7; bench/class_hold_statements_2026-10-20.md §2, whose druid row is superseded by that file).
• Offline state at this landing: bash tests/run.sh → 41 steps, 0 failed; luacheck → 2 warnings / 0 errors in 83 files (the standing floor; both warnings are pre-existing, in files this wave never touched). Every behaviour above is verified offline only.

Neutral actions, family scope, HUD ids, progression and contracts (2026-10-20, verified offline, UNCONFIRMED in-client)

• The NEUTRAL action — the engine half. A class publishes the action keys that are legitimate but never asked for, and the engine stops grading them as mistakes. The publish is ST.<CLASS>.practice_neutral_actions — flat, or one bucket per family (the bucket for the running st.family; a family the class published no bucket for answers nil, never a guess), read through one accessor, H.practice_neutral_actions(class_key) (nil-safe, nil for an unknown class, and it hands out a copy of the published set). H.rotation_run_decide gained a neutral outcome: the press records no attempt — no points, no streak, no mastery tick, no XP — and it is not a mis-pick. The panel shows NEUTRAL in the popup and the status line, and the run summary reports the neutral count only when it is above zero, so a class with no published set reads exactly as before. A HOLD still beats the neutral set (a cast during an explicit hold stays an illegal cast, F) and the rotation action beats it too (the expected action is graded first, so a rotation press still hits). The neutral outcome lives in the rotation scenario; a drill-list press is graded by the class's own drill grader and is not neutral there (bench/practice_neutral_actions_2026-10-20.md).
• The per-class sets. All six classes publish one today: the paladin (family-scoped — sotc, aw, hpot, abacus, stop, plus the common abilities holy_light, flash_of_light, lay_on_hands, divine_shield, bop, food on Era, TBC, WotLK and Forever, and the same ten keys WITHOUT sotc on Cata, Mists and Retail — the Seal of the Crusader went away in 4.0, so the class publishes it on the classic line only and a press of that key on the 4.0+ clients is a mis-pick, the rogue's Shiv treatment on the same clients), the warrior (family-scoped — sprint, demoralizing_shout, healthstone, flask, food on every family, battle_shout on WotLK, cleave on Forever), the rogue (family-scoped — sprint, evasion, kick, healthstone, food on every family, coldblood on Era and WotLK, shiv on WotLK and Forever) and the shaman (flat — healing_wave, chain_heal, water_shield, healing_potion, food). The hunter and the druid publish family-scoped buckets as well — Era, TBC, WotLK and Forever for both, the hunter's without raptor on WotLK and Forever and the druid's without rake on WotLK and Forever. The four classes that publish no bucket for Cata, Mists or Retail — the warrior, the rogue, the hunter and the druid — offer no common ability at all there (the paladin and the shaman still answer every family: the paladin with its ten-key 4.0+ bucket, the shaman with its flat set) (bench/class_neutral_family_coverage_2026-10-20.md §1.2/§5; bench/final_gate_fixes_2026-10-20.md §2). The common abilities needed a key first, so each of the six classes publishes a practice-panel ability list (ST.<CLASS>.practice_common_actions — the English spell name first, the cited id ladder second, family-correct: the name on TBC, the numeric id elsewhere) and appends it to the class's own guide slot list inside the practice block of post_init, after the guide frame built its icons — so the live next-action guide is untouched while the practice panel's action row offers them with their own icons. An ability the running client does not have is dropped (unknown-hides): the panel never offers a button the client cannot cast. A rotation action may never be published as neutral, and the panel vocabulary stays a closed list — every key is classified rotation / neutral / intentional mis-pick (bench/class_neutral_actions_2026-10-20.md; bench/class_common_actions_2026-10-20.md).
• The shaman's practice surface is family-scoped now. The rotation table's per-row families list was read by nothing at runtime and there was no accessor, so the panel fell back to the guide-slot vocabulary and offered every row on every client — a TBC client was offered the WotLK-only Call of the Elements and Bloodlust rows and the Era-only Molten Blast, Lava Lash, Chain Lightning, Windfury Totem and Grace of Air rows, so pressing one could only ever grade wrong. ST.SHAMAN.practice_rotation_for(family) now answers the family's rows in the table's own order; the argument-free call answers the client's family plus the running build's token, so the SoD build's own row rides the Era slice, and an unknown family answers an empty slice. ST.SHAMAN.family_guide_slots cuts the guide's slot set from the same families list, and ST.SHAMAN.apply_family_slots() narrows the published guide.slots to the running family's set (called from post_init after the guide frame and the HUD are built, and again when the SoD build becomes known). The per-family slices: Era 13 rows, TBC 13, WotLK 13, Forever 7, and the SoD build's own slice 14 (the Era slice plus its Lightning Shield row). ST.SHAMAN.PRACTICE_PREPULLS has no consumer — the class ships no pre-pull drill — and the class file now says so (bench/shaman_family_scope_2026-10-20.md).
• The hunter's HUD slots carry the running family's own rank. One owner for both surfaces: SPELL_ID_BY_FAMILY — the HUD slot and the guide slot of a key read the same entry, so the two can never disagree about which spell the running client has; a family with no entry for a key cannot have that spell, and the HUD's dead-slot filter drops the slot while the guide's slot set does not declare it (the Steady Shot keeps its own owner). The runtime re-point now re-points every spell slot with the client's own rank, not only the Steady and the Aimed ones. The per-family HUD slot counts: Era base 6, Era/SoD 7, TBC 9, WotLK 9, Forever 7, Cata 9, Mists 8, Retail 9 — the Forever row is seven slots, because the Forever client has neither a Steady Shot nor a Kill Command and both slots are dropped (bench/hunter_hud_family_2026-10-20.md).
• The progression surface, six fixes. (1) The daily-cap floor — the cap is max(3 x the median of the drill's HIT points, its best hit in the window), taken over hits only, so a miss (which earns nothing) no longer drags the cap down; without the floor a drill below 50% accuracy had a cap of 0 and swallowed every award for good. (2) The cap notice — the swallow is now reported: "Practice XP: this drill has reached today's XP cap (3x its median hit XP); it earns again after the next day.", once per drill per day, through the addon's own printer and not the badge-gated announcer, so a progression dead end is never silent. (3) The account scope — the session end evaluates the class scope and the account scope (P.evaluate_unlocks_both), so an account-scoped achievement can fire at all; the unlock flags live in the scope they were earned in. (4) The streak readout — the stats summary and the stats frame's summary line now carry the daily streak next to the lifetime best ("Best streak %d | Daily streak %d"); it was computed and never displayed. (5) The reset-arm expiry — the two-step reset's armed state now expires after 15 seconds, checked lazily at every decision point, so a stale "Confirm: reset all practice progress" can never wipe the profile minutes later. (6) The NO_DAY sentinel — the history day is read signed, so the "no date recorded" marker (-1) survives the repair and the export and renders as "-" instead of 1970-01-01, while a day of 0 (the epoch) stays 0. Alongside them: a saved blob a newer build owns is protected and reported ("this session's XP is not being saved"), and the logout flush reports the write, not the close (bench/progression_fixes_2026-10-20.md; bench/progression_ui_fixes_2026-10-20.md).
• The drill-list press is the pick. The press is written into every alias the class graders read (ctx.pick, ctx.pressed, ctx.action), so a class that publishes its own pick alias can no longer shadow the player's press. The block used to fill only an empty alias, and the hunter publishes ctx.pick (the live cast) and reads it first — a drill-list press made while a Steady Shot was casting was graded against the live cast instead of the press (measured: grade F for a correct press). A class's own pick still survives a press-less drill press (the hunter's shot drills judge the shot in flight), and the scenario path — a different call site — keeps the live cast as the pick. The contract is pinned against the six classes' real graders: each class is driven end to end through the drill-list path (the action press and the drill press), each answers a letter, and the pick the rotation wanted grades S; the negative half (no pick -> nil) is what makes the positive half mean something (bench/drill_pick_contract_2026-10-20.md).
• The pre-pull contract suite. NEW tests/prepull_contract_state.lua proves the contract between a class's pre-pull drill and that class's own pre-pull table over the eight published tables of the six classes: each is proved GRADED (the hunter's and the warrior's, with their late-band tables, and the rogue's) or CITATION-only (the druid's, the shaman's and the paladin's rows table) — a citation table has no reader anywhere in the addon tree and the comment above its declaration carries the marker, and a drill reads its own table and no other class's (coupling and cross-isolation). The paladin is the one class with both shapes: paladin_prepull grades the class's own per-order seal spec, and ST.PALADIN.practice_prepull_rows is the marked citation layer. The paladin's flags are now strict booleans on every row plus the needs-when-unread rule over both axes, and the druid's table gained the "no consumer — citations only" marker (bench/prepull_contract_2026-10-20.md).
• The documentation-truth corrections. A sibling lane re-audited README.md, TESTING.md and the operator's live script against the tree and corrected them: /st diag verbose does not exist — the handler matches only the empty argument, practice, design and diag, and /st diag is the verbose report; the rogue's Slice and Dice upkeep is armed, so the script no longer asks the operator to confirm its absence, and the rogue's Forever pre-pull is Blade Flurry 13877 @-2 with the builder chain {Saber Slash, Mutilate, Hemorrhage}; the Cata (40402) and Mists (50500) flavours are listed and explained (those two ship without the swing-engine library, so the bars are inert); the expected /st diag shape carries the ticks line; the paladin's utility layer is recorded as landed; and the usable-only trinket gate and the warrior's W-21 HUD mirrors got their operator steps. This changelog was audited in the same pass and is corrected here: the paladin pre-pull names, the warrior's drill count, the design-target count, the void correction notice and the two copy strings (desks/docstruth-audit/journal.md).
• Offline state at this landing: bash tests/run.sh → 43 steps, 0 failed; luacheck → 2 warnings / 0 errors in 85 files (the standing floor; both warnings are pre-existing, in compat.lua and main.lua, files this wave never touched). Every behaviour above is verified offline only — the in-client checklist in TESTING.md and the operator script in bench/live_test_script_2026-10-19.md are the owner acceptance pass.

Operator test-session wave (2026-10-21, verified offline, UNCONFIRMED in-client)

Found by the operator's first in-client session on Classic Era 1.15.9; the range fix and the design-mode work are verified offline, the in-client confirmation is still owed.

• Fixed — the range poll no longer calls a client-BLOCKED API. On Classic Era the addon raised [ADDON_ACTION_BLOCKED] AddOn 'SwedgeTimer' tried to call the protected function from the vendored LibRangeCheck-2.0 item-range checker (LibRangeCheck-2.0.lua:498, reached from main.lua:1797): the library resolves IsItemInRange to C_Item.IsItemInRange on every modern client (on the classic branches the legacy global is that function — Blizzard_DeprecatedItemScript/Deprecated_ItemScript.lua:29), whose call the client blocks for addon code and whose argument shape is an ItemInfo, not the legacy item id. The library now builds no item checkers and no spell checkers on any shipped family, so the melee/ranged ladder is one uniform thing everywhere — the interact-distance bands: melee 8 yd, ranged 28 yd (Tauren 6/25, Scourge 7/27) — and it calls zero client range APIs on Classic Era, TBC Anniversary, WotLK Titanforge, Retail and WoW Forever. The old per-family divergence (the legacy spell-range global exists in some client binaries and not in others) can no longer change the behaviour. Accepted coarsening, stated plainly: the range band is the client's interact-distance band, not the class's spell range, so the bar's range readout reads a band instead of the class's own reach. The accurate per-class range (the modern C_Spell path) is an open follow-up that needs one in-client confirmation; it is deliberately not shipped on a guess.
• Changed — design mode no longer fights the options panel. The design chrome now draws above the AceConfig options panel (strata FULLSCREEN_DIALOG, levels 110-150) instead of behind it, every full-screen design frame is explicitly mouse-transparent, and the per-element drag inputs sit below the panel, so the panel's title-bar drag can never be swallowed by design mode. Esc no longer ends the session while an options window is open, and closing the options panel leaves design mode active — you can now close the panel and keep designing.
• Added — a design control panel. While design mode is on, a small movable panel (no db key; a session state like design mode itself) offers Lock frames / Unlock frames (the overlays stay visible but frozen: no drags, no double-click reset) and Exit.
• Added — tools/deploy.sh. Copies the exact shipped file set to a client folder and verifies byte parity. The six Error loading SwedgeTimer/classes/*.lua warnings of this session were a hand copy that never carried the classes/ subdirectory — the files were in the repo and the TOC listed them, the deployed folder simply did not have them.
• Offline state at this landing: bash tests/run.sh → 43 steps, 0 failed; luacheck → 2 warnings / 0 errors in 85 files (the standing floor). tests/lrc_guard_state.lua → 170 checks; tests/designmode_state.lua → 655 checks. The new behaviour is verified offline only; in-client confirmation on Classic Era is owed for the range error, the design-mode layering and the new control panel.