Changelog

[v1.7.0] - 2026-10-20

Tests and tools: luacheck declared; gear EP audit tidy (tools only)

• .writ-suites.json declares a luacheck suite (luacheck .). Until now no suite ran it: the 44-suite gate on 2026-10-20 had no lint step for Lua at all. .luacheckrc now also excludes tools/wago_cache, the git-ignored downloads (1,905 third-party files, ~679,000 warnings). The 45 repo files are all clean.
• Tests/gear_ep_audit.lua: the gear filter tested INVTYPE_NON_EQUIP twice (an exact compare and a prefix match); only the prefix match remains (Peer Review 3988623c G4). The header now says that rows are scored with no options, so procs run at the flavour model's defaults, and what that misses (G5). G1's red-check was done with a real defect: removing TBC's whiteDPS turned gear-ep-tbc red on exactly Hand of Justice's extra-attack row, and restoring it turned it green.
• Tests/sources_order_check.py docstring: the "running build-sources-journal.py Cata" line it prints is the builder's own announcement before it calls the stand-in, not a rebuild. The check writes nothing.

Hand of Justice priced on TBC, Wrath and Mists; one whiteDPS definition (tools + scoring)

• Cause, traced in the client data: from TBC on, Hand of Justice's equip spell 15600 carries SpellAuraOptions.ProcChance = 100 (wago 2.5.6.68941, 3.4.3.58936, 4.4.2.60895, 5.5.4.68806), the "server decides" sentinel real_chance() refuses; Vanilla's client says 2. The item has no item_template rate either, so tools/build-use-effects.py found no rate and listed it unscored. The world DBs do carry it: spell_proc_event 15600 ppmRate = 1 in tbcmangos, wotlkmangos and mistsskyfire (desk itemdb-worlddb). The builder loaded that table already but read it only for on-use windows. It is now the LAST rate tier for an EQUIP proc with no other rate, so every row that already had a rate is unchanged.
• Data: UseEffects rebuilt pinned (TBC 2.5.6.68941, Wrath 3.4.3.58936, Mists 5.5.4.68806, Vanilla 1.15.9.69722 then split-seasonal.py Vanilla). HoJ ships extra_attack at ppm=1 on all three. The same tier gave a world-DB rate to other equip procs that had none, HoJ included: proc rows Vanilla 191 -> 205, TBC 237 -> 254, Wrath 317 -> 335, Mists 376 -> 397, with unscored falling by the same 14 / 17 / 18 / 21 (file headers; the build log names each as a spell ppm <item> line). Those items gain EP. On Vanilla that is 14 items, not HoJ, which already had ch=2. Cata has no world DB of its own (SPELL_PPM_DB is Mists-only on purpose, Peer Review a7919e37), so HoJ stays unscored on Cata.
• whiteDPS has ONE definition, written in the library's extra_attack bucket comment: the WEAPON DPS of one named raid two-hander, from the flavour's own core where it carries a DPS key and from WoWSims' database where it does not (Wrath), no attack-power term (a floor). Vanilla's 90 already was that (Corrupted Ashbringer 22691, DAMAGE_PER_SECOND=90 in our core); only its comment said "the wearer's white DPS", now corrected. TBC declares 120 (Gorehowl 28773, 119.86) and Mists 4186 (Starshatter 86140, 4185.97), both first-tier weapons like Wrath's Naxxramas-25 223. Cata declares none: it ships no extra_attack row. Pinned by the new Tests/tbc_scoring_spec.lua and a Tests/mists_scoring_spec.lua case. KNOWN COST: the tier rule is a first-tier raid weapon, and Vanilla is the exception: it keeps Naxxramas' 90, the last tier, because changing it would move live Vanilla scores (MELEE_HASTE_PCT and the weapon-skill default read the same number).

Mists mana procs priced for healers; gear EP audit reads ranged-AP twins (library + tools)

• LibItemDB-1.0.lua proc scorer: a mana-return proc (b = "mana") read the raw w[mp5] weight, so a flavour that DERIVES mp5 instead of weighting it scored every mana proc 0. Mists does exactly that (Scoring/Mists.lua prices mp5 as Spirit for healers), so the Idols of Wisdom, Price of Progress, Fathom-Brooch of the Tidewalker and the rest of 20 mana-proc rows scored nothing (gear-ep-mists run 20261020-102821.409-464835343). The mana branch now resolves its weight through effectWeight, like the buff branch, and its breakdown shows the derived factors. EP CHANGE on Mists: those procs now score for healer specs; every other spec and every other flavour is unchanged (only Mists declares a derived mp5). Pinned by a new Tests/libitemdb_spec.lua block. No operator ruling covers mana returns: pricing them rests on ItemDB's inference from the "healing has EP" ruling (directive 97bf22ed, which is about healing). If a ruling is wanted, the question is "should a mana-return proc have EP for healers?".
• Tests/gear_ep_audit.lua: on Cata and Mists a generic "+X attack power" effect ships as an AP row and an identical ranged-AP twin, and the twin is unpriced on purpose (the scales put a hunter's AP weight on melee AP; pricing both would pay a hunter twice, Scoring/Cata.lua "RANGED ATTACK POWER"). The audit failed all 54 / 52 twins; it now lists a twin (isRangedTwin: a melee-AP sibling on the same item with every field equal but the key and the amount) and still fails a lone ranged-AP row. Cata spell penetration is a documented zero there, as Scoring/Cata.lua already says. KNOWN COST: Cata 232012 Coren's Chilled Chromium Coaster ships melee AP 3576 beside ranged AP 4000 (Data/Cata/_core/UseEffects.lua:505; 71335, the same name, ships 4000 on both), so a Cata hunter is priced on 3576. Why the halves differ was not checked.
• Comments: Scoring/Mists.lua's socket block (Mists ships 17 cogwheel and 3 hydraulic gems) and mana-proc note, Scoring/Cata.lua's socket block, and the AP-pair note in the proc scorer (11817, 34472 and 34427 now ship the pair).
• tools/build-sets.py classify_proc read only the first stat aura of a set-bonus proc, the same defect build-use-effects.py had (Peer Review 9ddcdb05 N3). It now returns every stat aura and the caller writes one row per aura with the same rate fields; Wrath's equal spell-power / healing pair is settled by the shared settle_heal_twin, which moved from build-use-effects.py into itemdb_common so both builders use one copy. Every Sets.lua rebuilt pinned at its header build (Vanilla re-split): scored set procs TBC 35 -> 45, Wrath 39 -> 44, Cata 31 -> 36, Mists 34 -> 40, Forever 0 -> 0. The new rows are the second half of a two-stat proc -- on TBC, the ranged twin of an attack-power proc or the healing half of a "+damage and healing" proc (read from the TBC diff; the other versions' diffs were not read). The header now names ch|ppm and pm, which the rows carry (N4). EP CHANGE: set-bonus procs that buff two stats now price both halves as the scorer pairs them (AP max-paired; TBC healing and spell power priced per spec). Not checked in game.
• Tests/tarball_read_check.py gains an empty-regex case and a main()-level case (N5).

Tools only -- the gear EP audit's exceptions are tied to the scoring rules (Peer Review 3988623c, 737072573d5f, e750e48b)

• Tests/gear_ep_audit.lua section 2 now prices each on-use / proc ROW alone (the item's lists are swapped for that one row while it is scored), so a priced row no longer hides a dead one on the same item; a part counts only when (part.ep or 0) ~= 0. The summary line prints the row count and the section title says ROWS.
• Flavour-scoped zeros: Armor Ignored is excepted on TBC only, resilience on Cata and Mists only (TBC's scale weights resilience, so a TBC regression there now fails).
• On Cata and Mists the skipped keys are read from the rules actually registered (a copy, so the live rules are not mutated), not from a list kept in the audit.
• New SCORER_KEYS: a key priced by a registered lib.StatScorers function is never excepted, on any flavour -- WEAPON_DAMAGE (weaponDamage), AMMO_DAMAGE (ammo), and WEAPON_SPEED / MELEE_WEAPON_SPEED (speed, registered by Mists only). Before this, Mists excepted both speed keys as "not a stat" although Scoring/Mists.lua prices them, so a broken speed scorer would have passed.
• New failure class, "stale exception": an excepted key that is priced anyway fails, so a scorer that starts pricing an excepted key goes red until it is named in SCORER_KEYS. That is what keeps the hand-kept map honest.
• Red-checked 2026-10-20 by putting the defect in: with speed removed from SCORER_KEYS, gear-ep-mists failed on exactly WEAPON_SPEED and MELEE_WEAPON_SPEED as stale exceptions (run 20261020-132952.723-2215040992); restored, it passes (20261020-133101.345-1009074479). weaponDamage could not serve: no shipped Cata gear carries WEAPON_DAMAGE, so dropping it fails nothing.
• SCORER_KEYS gains sockets (every EMPTY_SOCKET_* key the sockets scorer reads), so it matches the four registered scorers; the weapon-speed exception reason no longer calls speed "not a stat", which is false on Mists (Peer Review 70bfa528).
• isRangedTwin ignores the amount on proc rows only; an on-use twin must match it too, since d and cd are all it carries beside the amount (Peer Review 9a3ffa52 N-A). gear-ep-cata 20261020-133316.577-1155008721, -tbc 20261020-133329.242-588795676 and -wrath 20261020-133554.172-941912331 pass with both changes.
• Comments: Scoring/Cata.lua names its one hydraulic-socket item, 60892, as hidden; the library's GEM_SUBCLASS_SOCKET says why one Cata/Mists map serves every flavour and that a Retail scoring file must make it per-flavour first.

School resistance under one key on Mists and Retail; Forever-only stats listed (tools + data)

• Gear EP audit finding: Mists and Retail cores carried a school resistance twice on one row -- ITEM_MOD_FIRE_RESISTANCE_SHORT=60 beside RESISTANCE2_NAME=60 -- so GetStats reported it twice and anything summing resistances doubled it. The builders were already right (build-core.py collapse_resist, build-core-wago.py emits RESISTANCEn_NAME only, and the Cata core was clean); the shipped data predated them.
• New tools/build-resist-collapse.py: applies build-core.py's own collapse_resist to every row of a version's _core and _seasonal class files in place, reads no DB2, and calls validate_version unconditionally. Needed because there is no Mists walk on this box to replay. Mists: 699 rows (948 duplicate keys dropped); Retail: 542 rows (671); Vanilla: 1 row (_seasonal/Armor.lua 213299). A second run changes nothing. No re-stamp: the class files keep their walk build.
• Tests/equip_resist_check.py: no core row on any version may carry ITEM_MOD_*_RESISTANCE*_SHORT. Red-checked against the real defect before the fix (Mists 948, Retail 671, Vanilla 1).
• Forever's EXPERTISE, ARMOR_PENETRATION, ITEM_MOD_ATTACK_POWER_VS_*_SHORT and ITEM_MOD_SPELL_DAMAGE_VS_*_SHORT stay as they are and unpriced: Classic Era has no such stats for Scoring/Vanilla.lua to price (expertise and armour penetration are Forever combat ratings 24/25 in wow-ui-source-forever PaperDollFrame.lua; the Era walk records no creature-type stat). Forever items' generic attack power beside the creature-type bonus is real (Beastslayer 11907: ItemSparse stat 38 next to stat 131). Tests/gear_ep_audit.lua lists them as known exceptions; Tests/forever_vocab_check.py pins all three behaviours. gear-ep-forever now passes with 0 unpriced (13 before).
• Not verified: that Forever's armour penetration is a flat armour-ignored value.
• No resistance number was lost by the collapse (self-audit 2026-10-20 R1): collapse_resist keeps RESISTANCEn_NAME and drops its twin without comparing them, so the new build-core.resist_conflicts names a single-school pair that disagrees (the all-schools key is left out: "+10 all" beside a larger fire value is extra fire), and build-resist-collapse.py prints each as CONFLICT. equip_resist_check.py pins the helper and reads every pre-collapse core row at commit a6bd125 through git: Mists 699, Retail 542 and Vanilla 1 rows carried both spellings, and every pair agreed (34/34). Those counts match the agent's own.

Cogwheel and hydraulic sockets priced; their gems no longer read as meta (tools + data + library)

• Gear EP audit finding: EMPTY_SOCKET_COGWHEEL (Cata 7 items, Mists 15) and EMPTY_SOCKET_HYDRAULIC (Mists 30, e.g. 86226 Claws of Shek'zeer) priced 0 for every spec, because the socket scorer's colour map had only R/Y/B/M/prismatic. Worse, tools/build-gems.py wrote every gem with no red/yellow/blue Pawn flag as meta M, so cogwheels and Sha-Touched gems were counted as candidates for META sockets and inflated Cata/Mists meta-socket scores.
• LibItemDB-1.0.lua: EMPTY_SOCKET_COGWHEEL = "C", EMPTY_SOCKET_HYDRAULIC = "H" and GEM_SUBCLASS_SOCKET = { [9] = "H", [10] = "C" } (around line 2666); a gem is matched to its socket by its core subclass, so the library is right even against an old Gems.lua.
• tools/build-gems.py: SOCKET_SUBCLASS per version (Cata {10: C}, Mists {9: H, 10: C}; per version because Retail numbers gem subclasses differently) and META_SUBCLASS = 6; colourless_kind() gives C/H/M from the subclass read out of the shipped _core/Gem.lua, and a colourless gem that fits no socket is skipped and listed on stderr instead of written as M (none did). Cata Gems.lua: 9 rows M->C (59477-59480, 59489, 59491, 59493, 59496, 68660). Mists: the same 9 plus 77540-77547 M->C and 89873, 89881, 89882 M->H; 60 meta, 17 cogwheel, 3 hydraulic. TBC and Wrath rebuilt byte-identical. Rebuilt through itemdb-build (the builder takes no --wago-build; its only versioned input is the cached Pawn file).
• Cata and Mists meta-socket scores DROP: the meta socket's best gem can no longer be a cogwheel.
• 98056 (Mists, subclass 9) has no gem row, and is now HIDDEN instead: it is an unobtainable duplicate of the Sha-Touched gem Crystallized Horror 89882 (no loot, vendor or quest source in SkyFire 5.4.8, absent from AllTheThings, same enchant 5115), so tools/curated_hidden.py lists it and Data/Mists/_core/Hidden.lua carries it. Search and the gem rankings no longer offer it.
• Pinned in Tests/gems_socket_check.py (new, declared in .writ-suites.json, 21 assertions; red-checked by emptying SOCKET_SUBCLASS). Not verified: in game, and the shipped-file half of that check was not run against the old Gems.lua.

Stacking proc windows priced: Blackened Naaru Sliver, Death Knight's Anguish (tools + data + library)

• Gear EP audit finding: Blackened Naaru Sliver (34427) shipped d=0 st=1 on TBC, Wrath, Cata and Mists and priced 0. Its equip spell 45355 (10% on hit, 45 s cooldown) triggers 45040 Battle Trance (20 s, 100% on hit), which triggers 45041 Combat Insight (+44 AP and +44 ranged AP, CumulativeAura 10, no timer of its own). proc_effect took the inner spell's 0 duration and the outer spell's stack cap. wowsims/wotlk melee_trinkets.go models the same 20 s, 10-stack window.
• tools/build-use-effects.py proc_window(): taken only when the middle spell has a duration and a 100% chance with the outer proc's scope, and the inner spell has no timer and a stack cap of 2 or more. One row per stat (both AP and ranged AP now ship), carrying the new ws field.
• LibItemDB-1.0.lua: averageStacks() (a port of the builder's average_stacks, same 0.05 s step); scoreProcEffects prices a ws > 1 row as m x averageStacks(trigger rate, d, ws) x uptime x weight, the trigger rate being the row's own (hitsPerSec). No world DB has a rate for the window, so none is baked in. README documents ws under GetProcEffect; it ships in MINOR 46, which is unreleased.
• The same shape moved Death Knight's Anguish (38212, +15 crit rating a stack) from LoadUnscoredProcs to priced on Wrath, Cata and Mists. Nothing else moved in those four. Each UseEffects rebuilt pinned at its header's build (TBC 2.5.6.68941, Wrath 3.4.3.58936, Cata 4.4.2.60895, Mists 5.5.4.68806). Vanilla and Forever UseEffects were rebuilt pinned later, with the one-row-per-aura change below; neither gained a ws row (Peer Review 6b47813f item 3). That no Vanilla or Forever proc has the two-hop window shape is the likely reason, not verified.
• Pinned in Tests/use_window_check.py (red-checked by putting the defect back, which failed on the exact old row) and in a 7-test Tests/libitemdb_spec.lua block, which also pins the capped averageStacks(0.5, 20, 10) = 4.791521286738454 (the builder's own figure, held in both ports) and the resulting ~187 EP (44 x 4.7915 x 20/45 x 2). The uncapped bound is ~196 EP (the spec's ws=100 fixture). gear-ep-tbc: 0 unpriced on-use/proc items.

One proc row per stat aura (tools + data)

• tools/build-use-effects.py proc_effect read only the first stat aura of a proc spell, so a spell buffing two stats shipped half. It now returns every stat aura and the main loop writes one row per aura, with the same rate fields. Shard of Contempt (34472) ships its attack power AND ranged attack power rows on TBC, Wrath, Cata and Mists; on Vanilla, 11817 gains its ranged twin and Felstriker (12590) gains HIT_PCT=100 beside CRIT_PCT=100 (3 s). TBC and Wrath ship the same two rows (Data/TBC/_core/UseEffects.lua:208, Data/Wrath/_core/UseEffects.lua:290) and price HIT_PCT linearly as hit rating (Scoring/TBC.lua:136, Scoring/Wrath.lua:129). KNOWN COST on Vanilla, TBC and Wrath: +100% hit is far past the hit cap and the price is not capped, so Felstriker's hit half is overpriced; by how much is not measured. Whether those TBC/Wrath rows are new in this rebuild was not checked.
• EP CHANGE for hunters (Peer Review 9ddcdb05 N1): 11817 on Vanilla, and 34472 / 34427 from TBC on, now ship the attack power AND ranged attack power halves of one proc. The proc scorer max-pairs the two (apPairEp), so a hunter's score takes the larger half, where before it took the melee half alone on 11817. That 11817 shipped melee AP only before is from the builder's first-aura bug, not from a diff (not verified); the size of the move was not measured.
• Wrath: a "+N spell power" spell carries a damage aura and an equal healing aura; the builder's new settle_heal_twin (Wrath only) drops the healing half when it equals the spell power and keeps only the excess. 47662 and 47671 (which had shipped the healing half alone under the first-aura bug) now ship ITEM_MOD_SPELL_POWER=234. Scoring/Wrath.lua prices any healing row left as spell power 1:1 for healer specs only (effectWeight gains only =).
• Every UseEffects file rebuilt pinned at its header build; Vanilla re-split with split-seasonal. gear-ep-wrath passes with 0 unpriced keys. Pinned in Tests/use_window_check.py and Tests/wrath_scoring_spec.lua.

Attack-speed on-use trinkets priced (tools + data + scoring)

• Kiss of the Spider (22954) shipped in LoadUnscoredProcs at 0 EP on Classic Era: its use spell 28866 is +20% melee (aura 138) and +20% ranged (aura 140) attack speed, and build-use-effects.py read neither. The aura -> key map is now itemdb_common.HASTE_AURA_KEYS, shared with build-aura-buffs.py (it was a private HASTE_KEY in each builder): 138 -> MELEE_HASTE_PCT, 140 -> RANGED_HASTE_PCT, and WoW Forever's aura 319 (the same spell's melee half there) -> MELEE_HASTE_PCT. Two auras naming one key add, through the shared haste_fold.
• Chance-on-hit haste procs: proc_effect returns every haste half of the spell (haste_by_key), and the main loop prices the half that speeds this weapon's own swing -- ranged haste on a ranged weapon (RANGED_INV), melee haste otherwise. Before, a ranged-haste proc was unscored while the same aura on-use was priced.
• Scoring/Vanilla.lua: DERIVED.MELEE_HASTE_PCT prices 1% as MODEL.whiteDPS / 100 main-hand weapon DPS, family = "melee", so hunters take only the ranged half (their scales already weight RANGED_HASTE_PCT). An approximation, stated in the comment; a DERIVED row is fixed at load, so a caller's opts.attack.weaponDPS does not reach it. Scoring/TBC.lua converts both halves to haste rating like Wrath/Cata/Mists.
• Charges-only on-use (Manual Crowd Pummeler 9449: 3 charges of +50% melee haste, no cooldown on Classic Era and TBC) is now priced as one use per CHARGE_USE_CD = 180 s, ItemDB's own figure, matching Forever's cooldown on the same item and Season of Discovery's 213407 casting the same spell. Positive charges only (negative charges mean a one-shot item that is deleted). KNOWN COST: three charges are three uses and the mace is kept once they are spent, so this prices a finite buff as if it were reusable without limit.
• Gear that casts a consumable's own spell is unscored (consumable copy in the build log): re-casting replaces the aura, so it adds nothing to the flask, elixir, food or drink a player already runs. A consumable is upkeep by Item.SubclassID (2 Elixir, 3 Flask, 5 Food & Drink) OR by its ItemEffect.SpellCategoryID (11 Item - Food, 59 Item - Drink, 79 Item - Potion (Non-Combat)) -- the category is needed because Classic Era's client files its consumables under subclass 0. Effect: 4696 Lapidis Tankard of Tidesippe (a drink's mana regen) is unscored on Vanilla, Forever and Mists (TBC, Wrath and Cata already left it unscored), and Forever 286541 Bael'dun Tankard likewise; Mists 81136 Cloak of Hidden Flasks (Flask of Falling Leaves) stays unscored.
• Rebuilt pinned, every row against HEAD: Vanilla +22954, +19343 Scrolls of Blinding Light, +9449 priced, 4696 unscored, re-split through split-seasonal.py; TBC +9449 (500 haste rating); Forever +22954 both halves, +9449, +19343, 4696 and 286541 unscored; Wrath +44096, an earlier fix never rebuilt into Wrath; Mists 4696 unscored; Cata unchanged. From TBC on, Kiss is +200 haste rating, which already shipped.
• Data/Mists/_core/AuraBuffs.lua (rebuilt at 5.5.4.68806 through the shared map): the five 10% melee-haste raid buffs (Cackling Howl, Serpent's Swiftness, Swiftblade's Cunning, Unholy Aura, Unleashed Rage) also carry RANGED_HASTE_PCT=10, which Mists' aura 342 grants. The ranged half is from the client data only, not seen in game. The generated header now names the haste keys each file's own rows carry (build-aura-buffs.py, Peer Review N2 and fd6f4b3b F3): Mists lists RANGED_HASTE_PCT, Wrath and Cata do not, because no row there has one. Wrath and Cata AuraBuffs were rebuilt pinned (3.4.3.58936, 4.4.2.60895) on 2026-10-19; git diff after that per-file header rebuild is +6 -6 on each, all in the header comment, every row identical to HEAD (Peer Review N3). aura_buffs_check now holds each modern header's key list equal to the *HASTE_PCT keys its rows carry (Peer Review N4; red-checked by putting RANGED_HASTE_PCT back into the Wrath header, which failed exactly that check, 1 of 74).
• Tests/libitemdb_spec.lua: four haste specs; Tests/use_window_check.py pins both halves of 22954, 9449's charge row, 4696 as a consumable copy, and every scored on-use whose buff lasts at least its cooldown (13143 Mark of the Dragon Lord on Cata/Mists, hidden 1912); the lib's no-premium comment now names 13143 instead of the unscored 81136.
• EP CHANGE in scoreUseEffects (LibItemDB-1.0.lua, Peer Review A6): a use buff that lasts at least its cooldown now scores with uptime 1 and premium 1 -- the use premium (the default, or a caller's premium) is dropped, because a buff that never falls off is a flat stat, not a burst. Shipped effect: 13143 Mark of the Dragon Lord scores lower on Cata and Mists than in v1.6.x (+44 spirit, d=1800 on cd=1800); 1912 Deprecated Reed Pipe also qualifies but is hidden on every flavour.
• Tests/use_window_check.py (Peer Review J1/J3/J4): one use_effects_blocks helper splits each UseEffects.lua for all three readers; a guard counts every [id]= row and {k= entry in the LoadUseEffects block against what ROW parsed, so a row the regex misses fails instead of vanishing (red-checked: narrowing ROW's m= to integers failed it on Vanilla, TBC, Wrath and Forever); and the full-uptime check is renamed on-use rows (served and hidden) with d >= cd, because it reads hidden rows on purpose (1912).
• Red-checked the Mists aura 342 ranged half (Peer Review D1): with 342 mapped to melee only and Mists AuraBuffs rebuilt, aura_buffs_check failed exactly Mists Unholy Aura 55610 is 10% melee AND 10% ranged haste (1 of 71); restored and rebuilt at 5.5.4.68806.
• Re-checked against the data on 2026-10-19 (Peer Review D2): use_window_check.py pins 13143 on Cata/Mists and 1912, which every flavour's Hidden.lua lists; Mists 81136 sits in LoadUnscoredProcs; and SoD 213407's ItemEffect row at 1.15.9.69722 casts spell 13494, the same spell as 9449's.
• Also: Data/Vanilla/_core/Effects.lua held two rows joined on one line (4423, 4424), so its header stated 231 rows and data_shrink_check reported a loss. Regenerated with build-effects.py at 1.15.9.69722 and re-split; it is byte-identical to HEAD again.

Developer test rings no longer ranked (data only)

• The "Indalamar's Ring of N stat" family (33986-33998, 34094, 34095, 34550, 34551; epic, ilvl 130, no source -- guildorder.com's 34095 page: "dungeon journal does not list a source") matched no build-hidden.py JUNK pattern, so rankings and BiS could serve them; 34095 ranked first by melee hit. New pattern r"^Indalamar's " -- measured over every version's enUS names, it catches only that developer's test items (the rings, "Trinket of Pwnage", "Super Hot", "Debuffer", "Holy Hand Grenade"). Not confirmed unobtainable by an official source.
• Hidden.lua rebuilt pinned at each file's stamped build: TBC +13 (4,730), Wrath +15 (6,320), Cata +20 (10,070), Mists +21 (13,458); the Indalamar's ids account for TBC +13, Wrath +15, Cata +15 and Mists +16. Vanilla, Forever and Retail ship no such names.
• SIDE EFFECT of the same rebuild, not of the pattern: Cata and Mists Hidden.lua each gained 5 AllTheThings rows (ATT count 3,665 -> 3,670 and 5,784 -> 5,789), so these items are no longer served there: 43670 Prophet's Enchanted Tiki, 43697 Artifact from the Nathrezim Homeworld, 43726 The Idle Crown of Anub'arak, 43821 Faceless One's Withered Brain, 43823 Head of Cyanigosa. Why ATT's signal moved was not checked; a newer installed AllTheThings is the likely cause, not verified.
• Tests/hidden_patterns_check.py (new, declared in .writ-suites.json): real test names hidden, near-miss real names served, every shipped Indalamar's id in Hidden.lua. Red-checked by breaking the pattern (6 of 23 failed).
• Forever carries the same family under another name: "Leafre's Ring of ..." (WoW Forever's counterpart of the Indalamar's rings). New pattern r"^Leafre's " in build-hidden.py; over Forever's enUS names it matches only those three rings. Forever Hidden.lua rebuilt at 1.60.1.69913, 1,664 -> 1,667. The three names are in hidden_patterns_check.py; red-checked by breaking the pattern (1 of 36 failed). Not confirmed unobtainable by an official source.

Hidden.lua records which AllTheThings it read (tools + data headers)

• Peer Review (thread f3669537, finding 2): the side effect above could not be explained from the files, because the header stamped the build and the removed-with-patch cutoff but not the ATT release. att_hidden now returns the source it read, and build-hidden.py stamps it as -- AllTheThings read: installed <toc ## Version> (or source cache <sha> / not read (<why>)). The next ATT update shows as a changed stamp, separate from a JUNK change.
• Every Hidden.lua rebuilt pinned at its stamped build so each carries the stamp (ATT 5.4.3). TBC, Wrath, Cata, Mists and Vanilla: stamp line only (Vanilla re-run through split-seasonal.py, 0 rows moved). Forever and Retail moved, by ATT rows only (the name and level counts did not change); that their last build read an older ATT is the likely cause, not verified, since nothing recorded it then: Forever 1,664 -> 1,663 (191481 Tabard of Mastery no longer hidden); Retail 13,923 -> 13,927 (+25 hidden, among them the same five 4367x-4382x items as Cata/Mists, 144373-144379 and 151221-151264; -21 served again: 185784-185785, 250293, 17 ids in 255473-255491, 268551). Not checked item by item against the game.
• Forever's drop is now confirmed and fixed (Peer Review fd6f4b3b): installed ATT 5.4.3's db/Camelot no longer mentions 191481 at all, and that entry was its only Forever hidden flag. The Tabard of Mastery is a Season of Mastery reward (ATT u=16) and already hidden on Vanilla, Wrath, Cata and Mists. That it is unobtainable on Forever rests on the older ATT build having flagged it there (HEAD's Forever header counted it as ATT); it was not checked in game. It is now in tools/curated_hidden.py MANUAL_HIDDEN. Forever Hidden.lua rebuilt pinned at 1.60.1.69913: 1,664 again (5 ATT, 5 hand-curated); data_shrink_check and hidden_patterns_check (33/33 at the time) green.
• tools/build-hidden.py now logs every ATT id it does not count, split into "not in this version's names" and "already hidden by name or RequiredLevel" (Peer Review 6b47813f item 5). Forever's sixth ATT id is 4116, which Forever does not ship; that rebuild was row-identical (1,664). The shipped Forever Hidden.lua now holds 1,667 (1,639 by name): a later rebuild added the three Leafre's test rings (274978, 276765, 285326), hidden by name (Peer Review ca334b59 F3).
• Tests/hidden_patterns_check.py's data half widened from the Indalamar's family to every name ANY JUNK pattern matches, on every version with a Hidden.lua (33 checks). Red-checked with a real defect, an added ^Linen Cloth$ pattern no file was rebuilt with: 7 of 33 failed. The three Leafre's names (above, added later) took the check to 36.
• Self-audit: a non-zero att_extract.lua exit is now stamped (, extract exit N) rather than read as a clean ATT read. The ATT cache paths were spelled in three files (build-hidden.py, build-item-sources.py, fetch-att.py); the two builders now import ATT_CACHE_DIR / ATT_CACHE_META from itemdb_common, and Tests/itemdb_common_check.py holds the import-free fetch-att.py to them (red-checked: renaming the shared pin file fails it).

Seasonal test items hidden again on Classic Era (data + tools)

• Peer Review (thread 502f816e): build-hidden.py and hidden_patterns_check.py both called read_names(version, "enUS") without include_seasonal, so neither saw the names split-seasonal.py had moved to Data/Vanilla/_seasonal/enUS/Names.lua. The shipped Vanilla Hidden.lua had been built after a split, so seasonal test and placeholder items (209526 TEST Darkwater Talwar, 213391 UNUSED Electrified Mithril Gauntlets, 215235 2hander Animation as 1hander test copy, ...) were served on Season of Discovery and Fresh Anniversary realms. Both now pass include_seasonal=True, so the builder's output no longer depends on whether it runs before or after the split.
• Data/Vanilla/_core/Hidden.lua rebuilt pinned at 1.15.9.69722: 2,110 -> 2,137 (+25 by name, +2 by RequiredLevel > 60; ATT unchanged at 417). Red-check: the widened check failed on exactly these rows before the rebuild, then passed 33/33 after it.
• Peer Review (502f816e P1): the same defect was in ten more builders. Tests/seasonal_ids_check.py already required include_seasonal=True, but judged only builders some _seasonal overlay header names -- and a builder that never saw a seasonal id never made an overlay, so it was never judged. The check now judges every tools/build-*.py; seven are EXEMPT with a written reason (build-sources, -imenso, -journal, -wowsims, build-bis, build-bis-classic, build-rep-loot: their tables are in split-seasonal's skip set, so a seasonal row would load on Era unguarded). Red-checked by dropping build-hidden's flag: the check named build-hidden.py:363. build-expansions.py is EXEMPT too, found by the self-audit after I had first "fixed" it: Expansion.lua is grouped ([-1] = "id,id,..."), not one [id]= row per item, so split-seasonal cannot partition it, and reading seasonal ids put 6,522 of them in the BASE file as expansion -1 "not known", loading on Era. Reverted and rebuilt (17,605, all expansion 0).
• Peer Review (ab446d2d P1): the exemptions are now checked, not just written. split-seasonal's skip set is the module constant SKIP (Expansion added: it is keyed by expansion, not item id). seasonal_ids_check now asserts three things. Every table an EXEMPT builder writes in a split version, read from the _core headers, is in SKIP. No EXEMPT builder passes include_seasonal, which is the guard that was missing for the Expansion regression. And every table split does partition is [id]= or ["key"]= rows at column 0. Building it found two false reasons: build-sources-imenso and -wowsims write only Forever (which has no _seasonal), not skip-set tables. Their reasons now say so, and the check holds them to it. Red runs: with RepLoot dropped from SKIP, the check named RepLoot. Re-adding the Expansion flag and rebuilding went red on all three new pins, the shape pin naming Expansion.lua:15 [-1] = "...". Reverted and rebuilt; split then moved 0. 48/48.
• Self-audit (a88dc308 F1): an EXEMPT builder calling shipped_ids() (always base + overlay) is now refused too. Red-checked by switching build-expansions.py:75 to it.
• Peer Review (a88dc308 P2): core_rows is now a judged gate beside core_ids and read_names. It has the same include_seasonal=False default, and nothing was looking at it. The gate also reads a positional flag (core_ids(v, True)), per Peer Review N1. The first run went red on build-tokens.py:224 and on build-core-wago.py:398/:588 and build-core-wowsims.py:364. build-tokens now passes the flag, since TokenRewards is partitioned. Vanilla was rebuilt and is unchanged (87 tokens, 250 rewards); split moved 0. The other three calls sit in verify/report functions (read_core, verify_structure, verify_armor). Those are exempted by function, not by builder, in REPORT_ONLY, and each is held to making no write call. derive-core-maps.py is not judged: it writes no Data file.
• Peer Review (a88dc308 F2): two pins that the split moves only item rows. First, every Load table in a partitioned base file whose keys are mostly not shipped item ids must be named in NON_ITEM_TABLES, and all its keys must sit below THRESHOLD. Today that is only Reagents.lua's skill-line table, which split-seasonal's own docstring said stayed in the base "by position, not design". Second, no base TokenRewards row may name a seasonal-only reward, because split moves a row by its token key. Red runs: unlisting Reagents#2 named it, and adding 202254 to token 5824's rewards named TokenRewards.lua:15. Both reverted. Self-audit: the table-opener regex first knew only lib:LoadX({, so the 13 lib:LoadCore(n, { class files were never judged. Now it also matches lib:LoadX(<args>, {, and a pin refuses any opener it does not recognise. Red-checked with the old regex: all 13 class files named. 59/59.
• Peer Review (a88dc308 L1, L2). The no-write pin now also knows json dump, the shutil copy family, move/rename/unlink/mkdir/touch, and print(file=...) to anything but stdout/stderr. Bare copy and replace are left out, because dict.copy and str.replace share the names. An item-keyed table may no longer carry an unshipped key >= THRESHOLD, since split would move that row by a meaningless key. Red runs: a planted write_text in read_core named write_text@408, and a planted [299999] row in Effects.lua named Effects.lua#1. Both reverted. 59/59.
• Self-audit 6633cdd6 F4, which Peer Review agreed was latent: the REPORT_ONLY exemption now also checks every place the function is called. A call passes when it sits in another report-only or verify* function, when its result is thrown away (main's verify_structure(v, b)), or when it is inside an if whose body ends in sys.exit (main's if args.verify: branch for verify_armor). Any other call is named, because the base-only answer could then reach a write. The two allowances based on names or code shape are checked too. A verify* caller's whole body must make no write, and an allowed exit branch must make no write before its sys.exit. Red runs: _planted = verify_structure(v, b) in main named main@633, and a write_text planted in build-core-wowsims' --verify branch named exit branch@477:write_text@482. Both reverted. 68/68.
• Peer Review (4797cfd8 W1-W3). Every no-write pin uses a write set worked out from the code, not a fixed list. Any function in itemdb_common.py or the builder whose body writes, directly or through another such function, counts as a write. That brings in ItemDB's own Data writers, write_names, write_source and write_spell_schools, which the old list missed. open() now counts only in a write mode, so p.open("rb") is a read. Writes to the download cache are excluded when the target is CACHE_DIR, or a variable built from it or from a *_cache_path() helper. Without that, wago_csv's cache write made every DB2 reader look like a writer. New pins hold that exclusion honest: CACHE_DIR must sit under tools/, and every *_cache_path helper must build on it. shutil.copy and os.replace / os.rename / os.remove are now recognised, while Path(...).replace stays the known gap (W2). A comment records that the three report-only functions fill no container they are passed (W3). Red runs: write_source planted in verify_structure and in build-core-wowsims' --verify branch, plus shutil.copy in read_core. All three were named, and the copy also turned read_core into a writer, which flagged its verify caller. All reverted. 72/72. Peer Review's follow-up note: extractall, urlretrieve and rmtree now count as writes too. Red run: an extractall planted in the --verify branch named extractall@482. Reverted.
• Fixed: build-consumable-groups, -droprates, -factions, -gems, -item-places (both its core_ids and the gathering-node read_names), -mounts, -races, -reagents, -spell-schools, -use-effects. Vanilla rebuilt with each, then split-seasonal.py: 1,259 seasonal rows into seven NEW overlays, now listed in ItemDB.toc -- _seasonal/Factions.lua 517, UseEffects.lua 481, SpellSchools.lua 178, Reagents.lua 31, Hidden.lua 27 (the 27 above, moved), Mounts.lua 18, Races.lua 7. So on SoD / Fresh realms, seasonal items now have their faction and race limits, on-use and proc effects, spell school, reagent use and mount flag, none of which they had before. ConsumableGroups, DropRates and ItemPlaces moved no rows; all three are keyed [id]= (read), so a seasonal row would have been partitioned. That their sources (WoWSims, CMaNGOS) carry no seasonal item is the likely reason, not verified. Base files unchanged in count (data_shrink_check: no base shrink). None of it is checked in game.
• hidden_patterns_check.py reads _seasonal/Hidden.lua as well as _core.

On-equip spell stats one low on Classic Era and TBC gear (data only)

• Operator, 2026-10-18: Hand of Justice (11815) is +20 Attack Power in game; the core shipped 19. Root cause: the Classic Era and TBC clients' GetItemStats returns an on-equip aura's raw EffectBasePoints and leaves out the EffectDieSides (1) the tooltip adds, and the walk records GetItemStats. Seasonal items carry DieSides 0, so their walk value was already right; the Mists and Forever walks already report the tooltip value (Drake Fang Talisman 19406: 56 there, 55 here).
• tools/itemdb_common.py: equip_aura_rows (raw base + DieSides of every on-equip APPLY_AURA) and equip_die_corrections, which raises a key by its auras' summed DieSides ONLY when the walk's value equals their summed raw base -- so a corrected core, a seasonal item and a client that already reports the tooltip value come back unchanged. EQUIP_DIE_KEYS (aura -> walk key) was derived from the data, not from names: AP 99/124, spell damage 13 (all schools and per school), healing 135, mp5 85, health regen 161, equip-spell school resistances 22, and TBC's aura-189 ratings (defence, dodge, parry, block, hit, spell hit, melee/ranged crit, spell crit). Armour via aura 22 matched 2 of 12 items and is left out.
• tools/build-equip-spell-fix.py (new; desk itemdb-build builder equip-spell-fix) applies it to the shipped class files in place, by item id, without replaying the walk. Keys raised, Vanilla at 1.15.9.69722: AP 596, melee AP 31, ranged AP 74, spell power 1,794, spell damage 378, healing 632, mp5 513, health regen 38, resistances 74. TBC at 2.5.6.68941: the same families (AP 1,487, spell power 2,315, spell damage 1,458, healing 1,181, mp5 1,180, ...) plus 1,544 rating keys. A second run changes nothing. build-core.py applies the same correction to every new walk.
• KNOWN: the 15 TBC items whose walk folds melee and spell crit into one ITEM_MOD_CRIT_RATING fail the summed-base proof and stay one low. Wrath's on-equip stats: see the next section.
• Tests/equip_spell_stats_check.py (new, declared in .writ-suites.json): nothing left at the raw base, 4,879 Vanilla and 9,413 TBC keys at base + DieSides, and pins for 11815 (+20, the operator's tooltip) and 19406's TBC hit 20 / dodge 12 / AP 56 against the Mists walk. It was red on 11815 before the fix.
• Peer Review 868b8260: the proof runs per row (a base and a seasonal row are proved separately; 0 ids carry both today, so no data moved), and parse_blob / apply_blob in itemdb_common are the one blob rewrite, apply_blob raising on a delta it could not apply.
• Peer Review de472a90 finding 1: EQUIP_DIE_KEYS' (13, None) row also caught aura 13 misc 1 (PHYSICAL damage), so TBC Bronze Band of Force 30804 -- 13/126 and 135/126 at 5 + 1 beside a 13/1 at 0 + 1 -- was lifted by 2 to SPELL_POWER 7; the tooltip is 6. (13, 1) now maps to SPELL_DAMAGE_DONE only, which is how GetItemStats reports it (Might of Cenarius 21189, whose only aura is 13/1 at 3 + 1: TBC walk 3, Mists walk 4). equip_die_keys is the lookup (an exact entry wins), and equip_die_overlift undoes what the old lookup lifted; the fix tool ran it on TBC (30804 7 -> 6, the only row) and Vanilla (none). The check pins 30804 and 21189 and asserts nothing is left over-lifted.

Wrath on-equip spell stats (data)

• Data/Wrath/_core shipped every on-equip aura stat EMPTY: Hand of Justice 11815 and Drake Fang Talisman 19406 had no stats at all. build-core-wago.py builds Wrath from ItemSparse, which holds only an item's native stats; attack power, spell power, healing, mp5, health regen, armour and ratings on Classic gear are on-equip auras (ItemEffect -> SpellEffect).
• build-core-wago.py now decodes them under the classic profile (aura_stats, item_auras): the amount is base + DieSides (the tooltip); attack power and TBC ratings become ItemSparse stat slots so fold_ratings reports them exactly as stored ones; Wrath ratings take one unified key per group. The TBC walk gate holds it: 0 unexplained over 30,031 items, with 229 rows explained as the walk's own raw-base omission (see the previous section).
• Wrath rebuilt at 3.4.3.58936: auras decoded on 9,237 items, of which 9,231 ship -- the other six (209528-209533) have no ItemSparse row, so nothing is written for them (the builder now prints them); 24,359 items with stats. 11815 AP 20; 19406 AP 56 / hit 20 / dodge 12; 21189 physical damage 4; 30804 spell power 6 plus physical damage 1 -- each equal to the Mists walk, which reports the physical part as spell damage (see the next section).
• WALK_DRIFT['TBC'] gains 34702: 2.5.5 ItemSparse armour 74, TBC walk 78, Wrath ItemSparse 78.
• Aura 123 misc 1 (ignore N armour) is decoded as ITEM_MOD_IGNORE_ARMOR for every version (see the spell-penetration section). No Wrath item carries it: 3.x converted those items to armour penetration rating, which Wrath ItemSparse already holds (904 items).
• Tests/core_wago_check.py compares the shipped core against ItemSparse PLUS the auras, asserts the decode found more than 5,000 items, and pins the five items above.
• Cata is built from the modern profile, whose ItemSparse already reports these stats (its 19406 carries AP 56); the aura decode is off there and Cata was not rebuilt.

The two on-equip aura decodes are held to agree (tools only, Peer Review de472a90 finding 5)

• itemdb_common.EQUIP_DIE_KEYS (the walk keys an aura is reported under) and build-core-wago.py's AURA_AP / AURA_KEY / RESIST_AURA / TARGET_RESIST_AURA / CR_BIT_STAT / 13-135 pairing (the Wrath decode) say the same thing in two shapes; the wago one folds ratings and attack power through ItemSparse stat slots, so it cannot simply read the other table.
• Tests/aura_decode_agree_check.py (new, declared in .writ-suites.json) decodes every EQUIP_DIE_KEYS entry as a one-aura item through the real sparse_stats and requires the keys to be ones that table names (or their EQUIP_STORED_KEY respelling) at equip_amount. What one side knows and the other deliberately does not is listed exactly: aura 22's magic schools (not decoded by the classic profile), 22's armour bit (not in the lift), and aura-189 bits 17-19 / 23-24 (no single-aura walk evidence). 42 checks; red on a planted AURA_KEY 161 respelling.

Spell penetration one high, and armour-ignore stored as spell penetration (Peer Review de472a90 finding 4)

• Spell penetration is aura 123 (MOD_TARGET_RESISTANCE) with a NEGATIVE base. The Classic Era and TBC GetItemStats reports it negated and, like the positive auras, without the die: walk -(base), tooltip -(base + DieSides), so every walk value was one HIGH. Verified on three wowhead tooltips, 2026-10-18: Field Marshal's Silk Spaulders 16444 walk 11, "by 10"; Doomcaller's Robes 21334 walk 21, "by 20"; Shady Dealer's Pantaloons 30898 walk 176, "ignore 175". equip_die_corrections could not see it: it summed the raw base (-11) against the walk's 11.
• itemdb_common.EQUIP_AURA_SIGN ({123: -1}) and equip_amount make the lift, the respell proof, the fix tool's derivation and build-core-wago.py's decode read the aura as positive. EQUIP_DIE_KEYS gains (123, None) -> SPELL_PENETRATION.
• Aura 123 misc 1 (PHYSICAL) is "your attacks ignore N of your opponent's armor", which the TBC walk also reports as SPELL_PENETRATION, so casters' scales priced a rogue's pants. It is now stored as ITEM_MOD_IGNORE_ARMOR (ItemDB's own spelling: F:\Blizzard API Docs\GlobalStrings\enUS.lua has only ITEM_MOD_ARMOR_PENETRATION_RATING and ITEM_MOD_EXTRA_ARMOR, and a rating is a different stat -- Peer Review 6e9ed265 F2) through EQUIP_STORED_KEY, and physical_damage_renames / physical_damage_mixed are renamed stored_key_renames / stored_key_mixed because they now cover both.
• Data: TBC 192 rows lowered by 1 and 138 respelled; Vanilla 35 lowered (30 base, 5 seasonal; no Vanilla item carries misc 1). build-equip-stats.py emitted misc 1 as SPELL_PENETRATION too -- and GetStats fills a key the core lacks from EquipStats, so 30898 would have carried both keys after the respell -- it now emits ITEM_MOD_IGNORE_ARMOR; TBC EquipStats.lua rebuilt at 2.5.6.68941 (138 rows). Wrath rebuilt; the TBC gate's walk-raw-base explanations fell from 229 to 37 and SPELL_PENETRATION (54) and IGNORE_ARMOR (138) are exact.
• Scoring: Scoring/TBC.lua lists ITEM_MOD_IGNORE_ARMOR ("Armor Ignored") with no shipped weight. KNOWN COST: it prices 0. HawsJon's TBC ArmorPenetration mapped onto it gave Arms 1.1 EP per point (175 ignored ~= 430 AP), which I could not confirm is a per-flat-armour unit, so the mapping was reverted and BiSWeights.lua is unchanged. So Arms and Fury scoring prices Shady Dealer's Pantaloons 30898's 175 armour ignored at 0 until a weight with a confirmed unit exists.
• build-equip-stats.py decodes aura 123 through itemdb_common.equip_amount with the row's own EffectDieSides instead of a hard-coded base + 1 (Peer Review 6e9ed265 F5), so the sign rule has one copy. Vanilla and TBC EquipStats.lua rebuilt at their pins: no row moved.
• Pins: equip_spell_stats_check (51): 16444 / 21334 at 10 / 20 on Vanilla and TBC, 30898 IGNORE_ARMOR 175 and no spell penetration, the negative-base lift on 16444's real rows (11 -> -1), the respell on 30898's, and 30804's real rows as the live witness of a MIXED key (13/1 beside 13/126, counted by stored_key_mixed and left as walked by stored_key_renames, Peer Review 6e9ed265 F4); red on 8 before the data fix. core_wago_check (59): Wrath 16444 at 10. libitemdb_spec: a TBC fire mage's spell-penetration weight prices SPELL_PENETRATION 175 and not IGNORE_ARMOR 175.
• The TBC gate's walk-raw-base explanation is a value coincidence, so it now applies only to the 37 (item, key) pairs checked in at tools/walk_exceptions.json (build-core-wago.RAW_BASE); a new match fails the gate and shows as a diff of that file (Peer Review de472a90 F6). core_wago_check (61) holds the observed set equal to the file both ways; red on the empty list. It also refuses a pair written twice (the json loads into a set, which would hide it), and explain()'s docstring says the raw-base test runs first, so a mismatch another reason would cover must still be listed (Peer Review 2f81f49d N1, N2).
• WALK_DRIFT (the 12 TBC items whose armour the walk and every cached ItemSparse disagree on) moved out of build-core-wago.py into the same file, with its explanation, so both per-item exception lists live in one place in one form. verify() now reports the drift ids it used, and core_wago_check holds them equal to the file both ways, where it used to check only that the count did not exceed the list -- a listed item that stopped drifting went unnoticed. Red on a bogus listed id.
• aura_decode_agree_check (51) counts resilience bits 14-16 among the unproved rating bits, and sweeps every aura id 0..511 against 27 masks: anything that decodes must be an aura EQUIP_DIE_KEYS knows, and each any-misc entry must agree on every mask it decodes (162 pairs), not one representative (Peer Review 7fb1a54b). Red on a planted constant-free aura-200 branch and a planted (161, 4) respelling. No unexpected hit.
• aura_decode_agree_check (54) now also holds the aura-189 masks EQUIP_DIE_KEYS does not name (13 in the sweep): each may land only on the keys of the named masks it shares a bit with. Two exceptions are listed exactly: a single hit bit (32 or 64) folds to HIT_MELEE_RATING / HIT_RANGED_RATING, where the walk evidence (mask 96) is HIT_RATING. No TBC walk item carries either split hit key, so neither side has evidence against the other, and Scoring/TBC.lua aliases both onto HIT_RATING. The masks no named mask covers must be exactly the unproved rating bits. Red on CR_BIT_STAT bit 9 pointed at spell crit.
• That sweep only held single bits, so it now also decodes every aura-189 mask an on-equip spell really carries in the cached TBC and Wrath data (13 masks; the unnamed multi-bit ones are exactly 224 and 1792, pinned). 1792 (all three crits, Wrath item 867) folds to CRIT_RATING, a key neither named crit mask names; the TBC walk itself reports that fold for a 768 and a 1024 aura at one amount (16577, 29598, ...). It is listed by hand as FOLD_189 and asserted exactly, rather than allowed by calling the decode under test (Peer Review b7dc73db S1, 4b920cc8). Red on 1792 before it was listed. (54 -> 57 checks.)
• Item 34095 is HIT_RATING 100 in the TBC core and melee 200 / ranged 100 in Wrath because the item changed, not the decode: the test ring has no aura, and 2.5.5's row is melee 100 + ranged 100 ("Indalamar's Ring of 100 Hit Rating") where 3.4.3's is 200 / 100 ("...of 200 Hit Rating"). core_wago_check (63) pins both rows (Peer Review b7dc73db S2).

Physical damage no longer priced as spell damage (todo #8, Peer Review a93de54c)

• "+N physical damage" (aura 13 misc 1, Might of Cenarius 21189 "+4 damage") is reported by GetItemStats as ITEM_MOD_SPELL_DAMAGE_DONE, so the Vanilla and TBC walk cores stored it there and every scale priced a melee ring as spell damage. My Wrath rebuild above did the same, and Scoring/Wrath.lua aliases that key to spell power.
• The cores now store it as ITEM_MOD_PHYSICAL_DAMAGE_DONE_SHORT, the spelling the Forever walk already uses. itemdb_common.EQUIP_STORED_KEY is the one table (walk key -> stored key); stored_key_renames respells a key only when every aura reporting under it is a 13/1 aura and the value equals their summed base + DieSides, the same proof shape as the DieSides lift, so an item mixing physical and school damage under one number is left as walked. New stored_key_mixed counts those, and the fix tool and equip_spell_stats_check print the count: 0 in every Vanilla and TBC scope today (Peer Review 0c2a1e34 finding 3). New rename_blob raises on a missing or colliding key.
• Applied by build-equip-spell-fix.py (after the lift) and by build-core.py's lift_equip_dies: TBC 4 rows (21189, 21766, 24114, 32654; 30804 has no spell-damage key, its physical aura's raw base is 0), Vanilla 20 (21189 and 19 seasonal items). build-core-wago.py emits the stored key directly for every version and compares it in the TBC gate (4 exact, 30804's 1 explained as walk-raw-base, 0 unexplained); Wrath ships it on 5 items, and its spell-damage-only rows (97) are school damage.
• Scoring: Vanilla, TBC and Wrath list it in weightStats ("Physical Damage") with no shipped weight, so it prices 0 and no breakdown calls it unknown. Its value scales with weapon speed, so a flat weight would be wrong; a custom scale can still set one. The stale "one legacy item" comment in Scoring/Wrath.lua is corrected.
• Mists (Peer Review 0c2a1e34 finding 4): an earlier line here said the Mists wago schema has no EffectBasePoints, so the proof could not run. That was wrong -- SpellEffect at 5.5.4.68806 carries it (desk itemdb-schema); the Mists core had simply been built before the rename existed. build-equip-spell-fix.py Mists respelled 5 rows (21189, 21766, 24114, 32654 and 30804's physical 1, which Wrath's DB2 decode also gives), matching the TBC core by item id; 185987's spell damage 20 is a real school aura and stays. It also lifted 4 seasonal ranged-AP rows (185985, 186069, 199722, 209803) by 1, to the TBC walk's and Wrath decode's value: Mists' walk reports aura 124 at the raw base, unlike the other auras. Scoring/Mists.lua lists "Physical Damage" in weightStats like the other three. Pins: equip_spell_stats_check now covers Mists (21189, 32654, 30804, 185987, 185985; red-checked by putting 21189's old spelling back), and mists_scoring_spec holds 21189's 4 at 0 with no unpriced row (red-checked by removing the vocabulary row). In-game tooltips not checked.
• Pins: equip_spell_stats_check (40): 21189 physical 4 and no spell damage on Vanilla and TBC, nothing left to respell, rename_blob, the proof on 21189's real rows (4 respelled, raw 3 not), and build-core.lift_equip_dies turning a raw walk record's 3 into physical 4 -- red-checked by removing the rename from lift_equip_dies. core_wago_check (58): Wrath 21189 physical 4 / no spell damage, 30804 spell power 6 / physical 1. wrath_scoring_spec: 21189's 4 scores 0 for a holy priest with no unpriced row, where the walk spelling scores 4.

Melee + ranged rating twins counted once in scoring (Peer Review de472a90 X1)

• One aura improving melee AND ranged crit (TBC aura 189 misc 768) is reported under both ITEM_MOD_CRIT_MELEE_RATING and ITEM_MOD_CRIT_RANGED_RATING, and every flavour's rules alias both onto ITEM_MOD_CRIT_RATING. scoreStats summed them, so Gloves of Holy Might 867 (14 and 14 for one +14 line) scored as 28 crit rating. At least 391 TBC core rows carry such a pair (Peer Review's lower bound from a text search, not counted by me).
• LibItemDB-1.0.lua settleTwins: crit, hit and haste melee/ranged pairs that alias to the same key now count once, by attack power's sign rule (apPairEp: same sign takes the larger, opposite signs sum). A lone side and the generic key are untouched. This LOWERS the EP of every such item -- a correction, not a regression.
• Tests/wrath_scoring_spec.lua: three specs (each twin family, the sign rule, one breakdown row). Red-checked: with the settleTwins call removed exactly these three fail (947 passed, 3 failed). A fourth pins an unequal pair (12 / 20) as ONE breakdown row at 20: rows are keyed by the alias (ITEM_MOD_CRIT_RATING), so the melee-or-ranged spelling settleTwins keeps (Peer Review a93de54c F7, now the side apPairEp chose) never reaches a label.
• KNOWN COST (Peer Review de472a90): an item with a melee-only crit aura AND a both-types crit aura reports melee greater than ranged; settling keeps only the larger, so the ranged-only part is dropped for a hunter. Whether any shipped item has that shape is not verified.

Hand of Justice's extra attack priced in EP (MINOR 45 -> 46)

• Operator, 2026-10-18: Hand of Justice (11815) showed only its Attack Power in Dibs' EP and nothing for the 2% extra attack. tools/build-use-effects.py priced an extra attack as one swing of the item's OWN weapon damage, so a non-weapon had nothing to price with and was filed in LoadUnscoredProcs. It now ships {b="extra_attack", m=<swings>} for any item with no weapon damage in the client data; a weapon with damage (Ironfoe) keeps its damage row. The Wrath heirloom Battleworn Thrash Blade (44096) is a weapon yet ships extra_attack (Data/Wrath/_core/UseEffects.lua:445), so the builder found no damage for it; that a scaling heirloom carries none in ItemSparse is the likely reason, not verified.
• LibItemDB-1.0.lua scoreProcEffects: new extra_attack bucket (breakdown key PROC_EXTRA_ATTACK, label "Extra attack proc"): extra swings per swing (rate / swing rate x m, i.e. ch% x m) x white DPS x the DPS_MAINHAND weight only, since extra attacks are melee. White DPS is opts.attack.weaponDPS, else the new model.whiteDPS. Scoring/Vanilla.lua declares 90, the figure its weapon-skill scorer already assumed (WS_DEFAULT_DPS now reads it, one constant). Arms warrior: 0.02 x 90 x 5.31 = 9.6 EP added to HoJ.
• Assumption, stated: only swings are counted as triggers and the extra attack's rage and on-next-swing interactions are not modelled, so this is a floor, not a sim result.
• Data: Data/Vanilla/_core/UseEffects.lua rebuilt at 1.15.9.69722 (HoJ is the only Vanilla item: 190 -> 191 proc rows, 334 -> 333 unscored); Data/Forever/_core/UseEffects.lua at 1.60.1.69913 (10 extra-attack items, 477 -> 467 unscored).
• Superseded by "Hand of Justice priced on TBC, Wrath and Mists" at the top of this version: the missing half on TBC and later was the RATE, and TBC and Mists now declare a whiteDPS. Vanilla's row carries ch=2; Wrath's 44096 row carries ppm=1.08.
• Wrath now declares whiteDPS = 223 in Scoring/Wrath.lua's fallback MODEL, so the heirloom Battleworn Thrash Blade (44096) prices its extra attack instead of scoring 0. 223 is the weapon DPS of Betrayer of Humanity (40384, Naxxramas-25 two-hander): (606 + 910) / 2 / 3.4, read from WoWSims WotLK assets/database/db.json. It leaves out attack power, so it is a floor.
• Tools: tools/tarball-read.py (desk itemdb-tarball) gained a grep:<path-end>:<regex> mode, because db.json is too large to print whole. The tool's command did not change. A bad regex or a spec with no : now prints one error line instead of a traceback (parse_grep, which main sends every grep: argument through), pinned by Tests/tarball_read_check.py (new, declared in .writ-suites.json). That check covers parse_grep only, so a main that compiled the pattern inline again would stay green.
• Tests/libitemdb_spec.lua: seven specs (the warrior value, the hunter using the main-hand and not the ranged weight, the paperdoll override, a caster at 0, the itemized row's factors, the main-hand swing rate, a set stage). Red-checked by disabling the bucket: exactly the six value specs failed.
• Peer Review 366810d1. E1: the share divided by hitsPerSec, the same rate that multiplies a swing-triggered row, so the two cancelled and a dual-wielder (more hits than main-hand swings) priced HoJ low. It now divides by opts.attack.mainSwingsPerSec, else model.mainSwingsPerSec, else hitsPerSec (default scores unchanged). F2: the model / rate resolution was written at four sites and the weaponDPS override at two; new local procContext serves all four (both GetItemScore / GetItemScoreBreakdown item and set paths, GetSetBonusEP).

[v1.6.0] - 2026-10-17

The first four sections below ship (a second Vanilla BiS provider; threat and armour reduction priced in EP -- both MINOR 45; on-use rating trinkets priced from TBC on; IsBiS returning a fresh list); the rest are under tools/ or Tests/, which the packager does not ship, plus .pkgmeta (packaging only) and README.md.

Vanilla: a second BiS provider, "bisclassic", through phase 6 (MINOR 44 -> 45)

• Data/Vanilla/_core/BiSClassic.lua registers lib:LoadBiS("bisclassic", ...): 623 sets for 9 classes, phases 0-6, including the fire / nature / shadow / frost / arcane resistance lists. The wowsims provider stops early (hunter at phase 1, mage and shadow priest at 1, rogue at 2, warlock at 0 + 2, warrior tank at 1; WoWSims master checked 2026-10-16), so nothing in the library answered "what is BiS in AQ40 / Naxx" for most classes (Dibs thread 000389037f53, the Badge of the Swarmguard report). wowsims stays the default: ItemDB.toc loads it first.
• Source: BestInSlotClassic (GPL-2.0, github.com/Warkdev/BestInSlotClassic) at commit 302ab4ecca0c, cached by the new tools/fetch-bisclassic.py (desk itemdb-fetch, source bisclassic) and read by the new tools/build-bis-classic.py. Facts extracted, files not redistributed -- the AtlasLootClassic terms. Its model is a per-class/spec priority ranking plus each item's phase; the builder mirrors its SearchBis (an item counts from its phase on, lowest priority wins). Choices the source leaves to its UI, made here: phase 0 is phase 1 without raid or world-boss items; a fractional phase (2.5) counts from the next whole phase; PvP-rank gear is left out; random-suffix rankings collapse to the bare item id.
• Weapons and races (Peer Review 5ba689c6 B1-B3; the first cut got all three wrong and built warrior fury phase 1-2 from the Orc main hand 17068 and the Human off hand 18832, a pairing no race's list gives). Now: ONE SET PER RACE GROUP (the races whose rankings are identical), the largest group on the plain token and the others suffixed with their race names (default_human, default_orc, hunter default_dwarf_nightelf); one-hander and two-hander priorities are never compared -- the source numbers them as separate lists -- so a spec its WEAPON_ICONS offers both setups gets the one-hand set on its token and the two-hander on <token>_2h (dropped when it ranks no two-hander); rankings are keyed by (item, OffHand), so an item the source ranks as both a main hand and an off hand (19363, 19104, 19170) keeps both. Tests/bis_classic_check.py gains fixtures for each, red-checked by restoring the merge and the item-only key (both went red, then green).
• KNOWN, reported to Dibs: Badge of the Swarmguard (21670) is BiS here for Enhancement Shaman (phases 5-6), not for hunters -- BestInSlotClassic ranks it 12th-13th among hunter trinkets, behind Drake Fang Talisman and Jom Gabbar. IsBiS(21670) now answers for the shaman.
• CONSUMER-VISIBLE through an unchanged API: IsBiS(itemID) with no source reads every loaded provider, so on Vanilla it now also returns bisclassic hits -- including the resistance lists. A consumer that shows "BiS for ..." from it without a source will show more. Pass "wowsims" to keep the old answer.
• Items restricted to one faction (the source's Faction field) are not filtered: a set is faction-agnostic, like wowsims'.
• split-seasonal.py skips BiSClassic like BiS (it is keyed by class, not item).
• Tests/bis_classic_check.py pins the selection rule on a synthetic ranking.
• Peer Review c9097ad3 B6-B8. New lib:GetBiSSetSpecs(classID [, source]): the sorted spec tokens a FIXED SET source holds (default: the default set source), so a consumer finds default_human or tank_2h without rebuilding the naming rule. GetBiSSpecs was the only lister and reads the EP weight scales, which carry none of those tokens; its README row now says so. Same MINOR 45, because 45 is not released yet. README IsBiS row: one hit per spec token per phase, so a warrior item in all three race groups' lists gives three a phase. build-bis-classic.py race_rankings evaluated only the races some row named, so a playable race no row named got no set at all. New PLAYABLE_RACES (the original-Classic race/class matrix): with an unrestricted row present every playable race is evaluated and an unnamed one gets the unrestricted rows, as the source's SearchBis gives it; without one, a missing race or a named race the class cannot be is refused. The real cache builds the same 623 sets. Four new fixtures, red-checked by disabling the unnamed-race path (the check went red).

Vanilla and TBC: threat reduction and target-armour reduction priced in EP (MINOR 45)

• Operator, 2026-10-16: "threat reduction is HUGE for DPS classes". Fetish of the Sand Reaver (21647, -70% threat for 20 s every 3 min), Eye of Diminution (23001, -35% for 20 s every 2 min), Badge of the Swarmguard (21670) and, on TBC, Icon of Unyielding Courage (28121, ignore 600 armour for 20 s every 2 min) were in LoadUnscoredProcs, scored 0, and Dibs' list hid them.
• tools/build-use-effects.py: the on-use path now emits THREAT_REDUCTION_PCT / SPELL_THREAT_REDUCTION_PCT (aura MOD_THREAT, sign flipped so a reduction is positive) and TARGET_ARMOR (aura 123 MOD_TARGET_RESISTANCE, negative, armour bit). New use_window: an on-use spell whose aura is a proc trigger (the Badge) emits the AVERAGE stack count over its window -- Poisson arrivals at the proc's CMaNGOS spell_proc_event ppm (26480: 15 per minute), capped by the triggered aura's CumulativeAura -- 661.1 armour, not the 1,200 cap. New pure helpers use_keys / average_stacks, pinned by the new Tests/use_window_check.py (18 assertions, red-checked by restoring the old sign test).
• Scoring/Vanilla.lua / Scoring/TBC.lua: derived rules for the three keys through the spec's crit weight (1% crit taken as 1% damage, 0.5% for spell crit). Armour: (A + K) / 100 points per 1%, Vanilla A = 3731, K = 5500 (WoWSims Classic level-63 boss and armour constant); TBC A = 7685 (Magtheridon), K = 10557.5, through crit rating. Threat 1:1 with damage is an UPPER BOUND, ItemDB's own and stated as such; no community weight exists for it.
• LibItemDB-1.0.lua effectWeight: a derived rule may carry exclude = { [classID] = { [spec] = true } }, read from lib.bisWeightsSpec. Threat reduction excludes every tank and healer spec. A scale not registered through LoadBiSWeights has no spec and is not excluded.
• Data: Data/Vanilla/_core/UseEffects.lua and Data/TBC/_core/UseEffects.lua rebuilt at the builds their siblings carry; the diff is exactly those items leaving the unscored list. The post-build split-seasonal.py added its partition marker line to Data/Vanilla/_core/ItemPlaces.lua and TokenRewards.lua, which had not been through it before; no row changed.
• Tests/libitemdb_spec.lua: seven specs on synthetic rows (armour and threat pricing, the tank / healer exclusion, an unregistered scale), red-checked by disabling the exclusion.
• Peer Review 917f21ea follow-ups. The exclusion list is ONE table, lib._noThreatSpecs in the core, which both Scoring/Vanilla.lua and Scoring/TBC.lua read (each carried its own copy). New specs: both flavours' threat rules hold that table; every spec it names is a spec the shipped Vanilla, TBC and Forever scales load (red-checked with feral_tnak); and no two specs in any flavour's shipped scale share one table, since lib.bisWeightsSpec is keyed by table and a shared one would carry the last spec's exclusion. The TBC [2].default and [5].default scales are byte-identical to holy and discipline, so excluding them is correct. Data/Forever/_core/UseEffects.lua rebuilt at 1.60.1.69913 (Forever loads Scoring/Vanilla.lua): 15 -> 17 on-use rows, 480 -> 478 unscored.
• Peer Review a7919e37. The Badge of the Swarmguard (21670) still scored 0 on Forever: use_window read its ppm from CMaNGOS only, and Forever has no CMaNGOS DB. It now falls back to the version's OWN wowsims ppm (by the on-use spell, then by the item; a per-hit chance is not a ppm and is skipped), never Vanilla's CMaNGOS rate, which build-sets.py removed as a defect. Forever rebuilt at 1.60.1.69913: 17 -> 18 on-use rows, 478 -> 477 unscored; the Badge prices at 481.3 armour (Vanilla, at CMaNGOS's 15 ppm: 661.1). Tests/use_window_check.py now also reads Forever's file (26 assertions). New spec, the reverse of the token check: every tank or healer scale in the Vanilla, TBC and Forever weights (Stamina weighted 0.5 or more, or healing 0.5 or more and above spell damage) is in lib._noThreatSpecs. Red-checked by dropping shaman tank from the list, and the shared-table spec by aliasing one Vanilla rogue scale: exactly those two failed. The _noThreatSpecs comment now names Forever.

TBC to Mists: on-use rating trinkets priced; the Badge scored on Wrath and Mists

• build-use-effects.py's use_keys had no branch for aura 189 (MOD_RATING), so every on-use buff that grants a combat rating (haste, crit, hit, dodge, armour penetration, mastery ...) fell through to the unscored list on every flavour from TBC on. It now decodes the mask through the shared itemdb_common.rating_keys(misc, version), the table build-effects.py and build-aura-buffs.py already use, so an on-use trinket and a food carrying the same bit name the same key. Rebuilt, each at the build its siblings carry: Wrath 115 -> 166 on-use rows (940 -> 889 unscored), Cata 189 -> 273 (1421 -> 1338), Mists 264 -> 399 (1954 -> 1820); TBC now ships 83 on-use rows and 601 unscored. Vanilla and Forever are unchanged (no aura 189 in a 1.x client).
• The Badge of the Swarmguard (21670) on Wrath: CMaNGOS wotlk has its ppm (15), but there its stacking aura is +28 armour penetration rating, not -200 target armour, so the missing rating branch was the whole cause. It prices at 92.6 armour penetration rating.
• On Mists: no CMaNGOS DB, and wowsims/mop does not model the Badge. use_window now also reads per-spell ppm from Mists' own SkyFire 5.4.8 world DB (mistsskyfire.sqlite, the same spell_proc_event columns), Mists only, so Cata never takes Mists' number. 26480 is ppm 3 there; the stack is +28 crit rating on that client, so the Badge prices at 21 crit rating. Meteorite Crystal (46051, SkyFire ppm 5) also gained a use window: 100 spirit.
• On Cata the Badge stays unscored: no world DB, and the cached wowsims/cata source has no Badge model (its only TBC trinket file is commented out).
• Peer Review cc49825a U1/U2: the build log now names the tier each use window took its ppm from (Forever: wowsims/forever spell 26480, ppm 10, 2.407 stacks), and use_window's docstring notes that the by-item wowsims tier keeps an item's first proc. Tests/use_window_check.py pins Forever's m=481.3, the Wrath and Mists Badge rows, and five use_keys aura-189 cases (33 assertions); red-checked by disabling the branch, exactly the three positive decodes failed.

IsBiS returns a fresh list (Peer Review cc49825a I1)

• lib:IsBiS(itemID) with no source returned the reverse index's own list, so a consumer that sorted or trimmed the result edited every later answer for that item until the index rebuilt. Both paths now build a new list (the hit records inside are still shared). New spec mutates one result and asserts the next call is unchanged; it failed (3 expected, 4 got) before the fix.

Tools only -- the armor-curve verify says what it skipped (Peer Review 3748bed0)

• My TBC self-audit said shields went unchecked because wago serves no ItemArmorShield at 2.5.6.68941. Wrong: no TBC shield carries walk armour (507 of them, the TBC walk's KNOWN COST), so verify_armor skipped every one at its want <= 0 test before the table mattered, and the run printed only exact / low / OVER -- a green line that read as if shields were covered. The verify now prints a "not checked" line: shields and cloth/leather/mail/plate with no walk armour (the latter named, up to 12). TBC at 2.5.6.68941: 507 shields and 6 body pieces skipped (4728, 5091, 19837, 33073-33075 -- ilvl-0 placeholders, a shirt-slot and a ranged-slot piece), 11,362 exact, 0 OVER. The comment that claimed the unchecked count stops a TBC run is corrected: it guards only a version whose walk carries shield armour (Vanilla, Forever, Mists, Retail). verify_armor now returns a third value, the skipped counts.
• A hidden item (test / placeholder, never served) that comes out OVER is excused, named and not failed; hidden items that match still count as exact. Forever's OVER row 16213 is "Test Quality Modifier Chest", whose armour is whatever a developer typed. (The first cut skipped every hidden row, which cut TBC's exact matches from 11,362 to 9,769 -- caught in the self-audit.) hidden_ids moved from build-core-wago.py into itemdb_common so both builders read Hidden.lua one way.
• Forever's other OVER row, 273646 "Half-Eaten Boots" (walk 43, curve 143), is now a listed ARMOR_EXCEPTIONS entry, printed by name on every run. The row is still compared: only an OVER result is excused, and an entry whose row is no longer OVER prints as STALE so it cannot go on excusing nothing. Its cause is NOT VERIFIED: wago's Forever ItemSparse at 1.60.1.69913 has no row for it (its Item row is there). A "-100 Armor" bonus would fit the gap exactly. Forever's verify now exits 0 (7,537 exact, 592 low, 0 OVER); Vanilla passes too (6,715 exact, 12 low, 0 OVER).
• Tests/armor_curve_check.py 13 -> 18 checks. The skipped-count check was written first and failed on the old two-value return.

Docs -- the changelog back under GitHub's release-body limit

• CHANGELOG.md was 127,330 characters, over GitHub's 125,000-character release-body limit, so the next release would have been rejected with HTTP 422. v1.1.3, v1.1.2 and v1.1.1 moved unchanged to CHANGELOG_ARCHIVE.md; checked section by section (32 before, 32 after, no word changed).
• markdownlint's line-length rule (MD013) stays off. Turning it on at 80 columns (Peer Review b4eb143b, M1) was tried and reverted on the operator's ruling, 2026-10-17: "there is only one line length rule, you need to move stuff from the changelog into the archive". The one length this repo keeps is the changelog's size, kept by archiving the oldest whole versions. The 80-column reflow tool written for M1 was removed with it.
• The markdownlint suite was shown to fail (with MD013 on it exited 1, so the npm shim passes the failure through), and it now lints CHANGELOG_ARCHIVE.md too (Peer Review b4eb143b, M2).
• CLAUDE.md: "merge- precedence" corrected to "merge-precedence", and the README rules' closing paragraph ("Write it for a dev...") moved out of the last nested bullet it had been read into.

Tools only -- derive-core-maps.py re-run on Vanilla, TBC and Mists (Peer Review d89d615b)

• First run of read_walk_core on itemdb_common.core_rows since the v1.4.1 refactor: every InventoryType maps to one token on all three versions, matching itemdb_common.INVTYPE. Every stat id's top key matches build-core-wago.STAT_KEY, or EXTRA_ARMOR / RESIST_STAT for ids 50-56, except Mists ids 11 and 40 (2 items each, no mapping; not investigated).
• The structural check now NAMES every item where wago and the walk disagree (the full list, not a truncated one), instead of only counting them. TBC's 2 are the WoW Tokens (122270, 122284), which the client reports as quality 8 with no subclass. Mists has 98 quality, 18 item-level and 7 subclass differences, where the shipped walk value is the client's own answer; nothing shipped changed.
• The comment claiming the true stat key "wins by orders of magnitude" is replaced with what was measured: about 3-4 to 1 for the primary stats, with ratings among the runner-ups on TBC/Mists.
• A stat id with five or fewer hits now lists its items. That named Mists ids 11 and 40: Test Sword Chest 13712, Test Sword Skill Trinket 33003 and Juno's Test Gem 34967, all in Data/Mists/_core/Hidden.lua. build-core-wago.py documents them as deliberately unmapped.
• build-core-wago.py gains verify_structure(): before it writes Wrath/Cata it reports where wago's quality, subclass, equip slot and item level disagree with the TBC and Mists walks, which verify() (stats only) never compared. Its docstring said the gate compared "every emitted key on every item", which read as covering structure; it now says the gate is stats only. Report-only, because no Wrath/Cata client is on hand to say which side is right; Tests/core_wago_check.py holds the counts exactly (TBC: the two WoW Tokens; Mists: 98 / 7 / 18), so a new drift goes red. Red-checked by reading item level from the wrong field: both counts failed.
• Forever also run: its ItemSparse carries no stat amounts, so nothing matches by value; its core is walk-built, so nothing shipped depends on that.
• Peer Review c53c25e2 nits: verify_structure's docstring names the equip slot it compares, and the header build-core-wago.py writes into Wrath/Cata files now says stats are gated and structure only reported. The shipped Data headers pick that up at their next rebuild.

Packaging -- .pkgmeta cleaned, and its checker declared

• .pkgmeta carried comment lines, which the no-comments rule forbids; removed. One was # curseforge-project-id: 1558208, which the packager never read: release.sh takes the id from -p or the TOC's ## X-Curse-Project-ID: 1558208 (in ItemDB.toc).
• Dropped the "*.sh" and "*.bat" ignore entries; no tracked file matches either.
• The harness's .pkgmeta checker is now the declared pkgmeta suite in .writ-suites.json. Its WARN [7] (no curseforge-project-id:) asked for a key release.sh does not read; reported to WoWAPITesting (inbox d446ea6d), fixed in harness f6e5fde, which reads the TOC id instead.

Docs -- GetTokenRewards named a GetSlot the library does not have (Dibs, inbox 65ade047)

• The comment above lib:GetTokenRewards and its README.md row told consumers to read each reward's slot with GetSlot. No such method exists, so following that raised a nil-method error. Both now name GetInfo's 5th return (equipLoc). Comment and docs only.

Tests -- harness pin 407bd26 -> f6e5fde

• Harness f5389cb runs a block's examples after its whole body, as busted does. That left Tests/pricewindow_spec.lua's file-level require("env.frames") + libs.load("LibAceGUIWidgets-1.0") in place before its two "without the widget library" examples, and both failed. The load now happens once, in the window block's first before_each. Suite 926/926.

[v1.5.0] - 2026-10-16

Added -- a Shaman tank weight scale (Dibs' contract 46c74258)

• GetBiSWeights(7, "tank") now resolves on Vanilla, TBC and Forever, in the default hawsjon source. HawsJon publishes no Shaman tank scale and no published Classic one was found, so tools/build-bis-weights.py's new DERIVED_SPECS copies HawsJon's own Protection Paladin scale -- the one published model of a mana-using shield tank whose threat comes from spell damage. Each generated BiSWeights.lua says in its header that [7].tank is a copy, not HawsJon's, and the builder refuses to copy once HawsJon publishes the spec itself.
• The copy drops DPS_OFFHAND (DERIVED_DROP). A paladin can never dual-wield, so that weight never applied to its own scale; a TBC shaman can (Dual Wield, 30798), and keeping it would let an off-hand weapon compete with a shield on a tank scale. Dropped on all three versions so they share one rule (Peer Review 4c1a17ee P3).
• PARRY_PCT is kept, verified from the client data rather than the talent's name: Enhancement talent 16268 ("Parry" on Era, "Spirit Weapons" on TBC) learns spell 18848, whose SpellEffect is the same effect (type 22) as the baseline Parry spell 3127, in both the 1.15.9.70003 and 2.5.6.68941 tables. Shield Specialization backs the block weights and Anticipation the dodge weight. Forever's tree was not checked. (The first cut of this entry and the builder comment said TBC had no Talents.lua and that Anticipation gives defense; both were wrong -- Peer Review 4c1a17ee P1/P2.)
• itemdb_common.header writes a blank header line as -- with no trailing space after it. Every builder's generated headers lose that whitespace the next time they are rebuilt.
• No scorer change: the Vanilla/TBC scoring rules carry no spec-keyed tables, and the per-class ScoreModel default applies. No MINOR bump -- the API is unchanged and Dibs detects the scale by GetBiSWeights returning non-nil.
• New Tests/derived_weights_check.py (declared in .writ-suites.json, red-checked by editing one weight in the shipped row, then by putting DPS_OFFHAND back) pins that each derived scale equals its source minus the dropped keys, that the header names it as a copy and each drop, and that [7].default is still elemental.

Added -- what a tier token is turned in for (MINOR 44, Dibs' contract 1c5f8e8d2faf)

• DB:GetTokenRewards(itemID) returns the gear a boss-dropped token is exchanged for, every class's together, as a fresh table sorted by item id, or nil for an item that is not a token (never an empty table). Dibs prices a token by a substitute item level and slot, and asked for a default derived from the rewards. lib:LoadTokenRewards replaces per token. MINOR 43 -> 44.
• New tools/build-tokens.py -> Data/<V>/_core/TokenRewards.lua for Vanilla (87 tokens), TBC (159), Wrath (226) and Mists (115), listed in those four TOCs. The client has no such table; the exchange is server-side, so it reads the same world DB build-droprates.py does: quest_template turn-ins (every Vanilla token) and npc_vendor/npc_vendor_template ExtendedCost against the client's ItemExtendedCost DB2 (TBC onward). A token is an item required by count 1, not itself equippable, dropped by an encounter in Sources.lua or DropRates.lua, and rewarding equippable gear in that version's _core.
• Three defects found while building it, each fixed and pinned by the new Tests/tokens_check.py (declared in .writ-suites.json, red-checked):
  • Head of Onyxia and Head of Nefarian were missing: quest 7490 takes the head and rewards nothing; its NextQuestInChain 7491 needs no item and rewards the rings. A reward-less turn-in is now followed down its talk-only chain (depth 3).
  • Chestguard of the Fallen Defender listed six Gladiator pieces beside its seven T4 chests. tbc-db prices Karynna's Season 1 stock (template 555) at the same ExtendedCost 1235 the T4 vendor uses, and the 2.5.6 client reads 1235 as one T4 chest token. A vendor now counts only when most of its cost rows resolve to a token (template 554 and npc 26089 fail that, at exactly half), and a reward carrying resilience is never a token reward. Every T4, T5, T6 and Sunwell row now matches AtlasLoot's own token table (AtlasLootClassic Data/Token.lua, BCC); 29753 and 30243 are pinned against it.
  • SkyFire's quest_template has no required-item columns, so Mists refused to build; the quest half now degrades to nothing there and the vendor half carries it.
• Not covered, stated rather than hidden: Forever, Cata and Retail ship no file: the pipeline has no world DB for any of the three. Mists carries the Wrath-era tokens only; why Pandaria's own tier tokens are absent was not investigated. Wrath's Trial of the Crusader trophies were not checked. Nothing ran in a live client.

Fixed -- the armor-curve verify hid its own failures (tools only, nothing ships)

• tools/build-core-wowsims.py verify_armor printed low and OVER rows from one list capped at 12. Low rows (items with bonus armor, expected, often hundreds) filled the cap first, so the Forever run of 2026-10-12 reported OVER 2 and named neither item. OVER rows are the failures, so they now print in full under their own heading; low rows keep the first-12 sample.
• That run's two OVER items are 16213 (walk 531, curve 541) and 273646 (walk 43, curve 143). 16213 is "Test Quality Modifier Chest", flagged hidden, and its Forever ItemSparse row (1.60.1.69913) carries QualityModifier=-10, which base_armor does not read; whether that field accounts for the 10-armor gap is not verified. 273646 ("Half-Eaten Boots", Forever-only) has no row in the cached Forever ItemSparse at that build, so the walk came from a newer client and the item cannot be inspected at the pin. Players are unaffected: Forever's core is walk-built, so its shipped armor is the client's own figure and this formula does not reach it. Rerun 2026-10-12 with the fix: both rows print under OVER:; Vanilla unchanged (6,715 exact, 12 low, 0 over).
• No shipped flavour carries curve-built armor today: every Data/*/_core/Armor.lua header (and Vanilla's _seasonal/Armor.lua) names build-core.py (a walk) or build-core-wago.py, and build-core-wowsims.py refuses to overwrite either. So neither OVER cause reaches a player.
• armor_curves treats ItemArmorShield as optional. wago answers HTTP 400 for it at TBC 2.5.6.68941 while serving the other three curve tables, and the TBC verify died there before checking a single piece. Shields are now "not computable" (base_armor returns None) and --verify counts them as unchecked, never as low (low means "carries bonus armor"). A 504 or any other error still raises. The TBC verify itself has not run yet: wago answered 504 at both 2.5.6.68941 and 2.5.5.67511 the same day.
• Peer Review (ad843502): unchecked rows first sat outside the total, so --verify exited 0 with every shield skipped, and a build at such a pin would have shipped every shield with no base armor. Now unchecked rows count toward the total and never toward a pass, so --verify exits 1 while any remain; and the new require_shield_curve makes the build refuse when the shield table is absent and any shield would ship. A wrong 400 then costs a rerun, never bad data.
• New Tests/armor_curve_check.py (declared in .writ-suites.json, 13 cases) pins that path, the 504 re-raise, the build refusal, the failing partial verify and the robe-reads-chest row. Red-checked three ways: the None guard replaced with an empty table (got 0), the old return restored (got (1, 1)), and the refusal short-circuited (got False).
• New repo desk tools: itemdb-verify-pinned (a builder's --verify at an exact --wago-build, so the check reads the build the shipped files carry -- wago.tools answered 504 for the installed 1.15.9.70003 while 1.15.9.69722 is cached) and itemdb-derive-maps (tools/derive-core-maps.py --version V --wago-build B, read-only; itemdb-build cannot pass its version flag).

[v1.4.1] - 2026-10-12

Fixed -- where-to-get places filed in the wrong zone (Dibs' DIBSREQ-IDB, Searing Whelp)

Dibs' report (inbox 2c393671f8ee): Dibs showed Small Flame Sac dropping from Searing Whelps in Thousand Needles on Classic Era. The whelps live in the Wyrmbog, Dustwallow Marsh.

• Root cause: tools/build-item-places.py files each spawn under the zone whose UiMapAssignment rectangle holds it. Those rectangles are map-image bounding boxes and overlap far past the real borders: every Searing Whelp (4324) spawn is inside the Barrens, Thousand Needles AND Dustwallow rectangles, and the tie-break ("the smallest rectangle") chose Thousand Needles. No rule over the rectangles alone gets it right (smallest, deepest-inside and nested were all tried), and the CMaNGOS creature_zone / gameobject_zone tables are empty.
• New tools/fetch-questie.py (desk itemdb-fetch, source questie) caches QuestieDB's (github.com/Questie/QuestieDB) NPC and object spawn tables for Classic, TBC, Wotlk, Cata, MoP and Forever at a pinned commit (cd4ac424, tools/wago_cache/questie.source.json), each file verified against its git blob id, and reads them: per entity, {AreaTable zone: [(x, y)]}. It shares fetch-att.py's GitHub helpers (resolve / fetch_one now take the repo's API, cache and raw roots, defaulting to AllTheThings').
• Placer.places takes that as a hint: a spawn inside several rectangles goes to the zone QuestieDB files the entity under (nearest recorded point when it lists several candidates), then the zone the entity's unambiguous spawns use, then the smallest rectangle. The AreaTable id becomes a UiMap through the client's own UiMapAssignment.AreaID (zone_areas, a SET of UiMaps per area, so an area on two zone maps hints both). The coordinates shipped are still the CMaNGOS spawn's; QuestieDB only picks the zone, and none of its data is shipped. QuestieDB carries no licence file (GitHub's licence API answers 404, recorded in questie.source.json); it is a build-time input only.
• A hint whose nearest QuestieDB point is more than HINT_MAX_YARDS (250) from the spawn gives way to the entity's clear spawns, but still beats the smallest rectangle. Measured on shipped spawns before choosing it: Vanilla, the same world as QuestieDB, median 0 yards, 99th percentile 49, max 4,644; Cata, whose places are SkyFire 5.4.8 spawns against QuestieDB's 4.x points, median 0, 90th 50, 99th 560. It moved four entities on Cata and Mists, none elsewhere: Coral Shark (5434) drops The Cape of Stranglethorn, Winterfall Runner (10916) drops Felwood, and the Firefin Snapper and Oily Blackmouth schools (180683 / 180682) drop Thousand Needles. Which side is right for those four is not verified.
• tools/fetch-questie.py refuses a line that opens a row ([<id>]) but is not one whole [id] = {...}, line, instead of skipping it, and looks the licence up before its already-pinned early return so a cache pinned offline does not keep "not reached".
• The builder reports which rule settled the overlapping spawns -- summed over the entities it ships, not every entity it placed (a vendor weighed for the zone spread and not chosen used to count) -- the distance spread above, any QuestieDB area that sits on more than one zone map (none on any flavour; its points would be measured in a frame that may not be theirs), and how many places changed zone against the file it replaces AND against git HEAD. Rebuilt every flavour QuestieDB covers; places that changed zone against HEAD, the cap included: Vanilla 775 of 4,725, TBC 916 of 6,510, Wrath 1,213 of 8,559, Cata 1,253 of 9,601, Mists 1,351 of 10,216. Vanilla's shipped overlapping spawns: 36,450 settled by QuestieDB, 344 by the entity's clear spawns, 96 by clear spawns over a far hint, 14 by a far hint, 33 by the smallest rectangle. Checked in the built file, not in a client: Searing Whelp to Dustwallow Marsh, Skeletal Warrior (48) and Gavin Gnarltree (225) from Deadwind Pass to Duskwood, Gruff Swiftbite (100) from Duskwood to Elwynn Forest, Mine Spider (43) to Elwynn Forest. Vendor choice shifts slightly where a vendor's zone changed (vendors are spread over zones), so a few translated names enter or leave PlaceNames.lua. Retail is unchanged: QuestieDB stops at Mists. Vanilla is pinned to 1.15.9.69722, the build its siblings carry.
• Tests/item_places_check.py: the hint against both old rules, nearest-point choice, an area on two zone maps, the distance cap both ways, the shipped-only report, a split row refused, the move report, and the real case on real data -- Searing Whelp places in 1445 only, and without the hint it still comes out 1441, so the check can fail. A missing cache fails it rather than skipping.

Changed -- Peer Review f4bc13b5 follow-ups

• Tests/itemdb_common_check.py's every-[id]= assertion now red-checks itself on every run: a scratch tree with one row core_rows cannot read must come out unequal.
• Data/Vanilla/_core/EquipStats.lua (hand-edited for the earlier red-check) is rebuilt by the builder at 1.15.9.69722 and re-split, rather than restored by hand.
• New desk tool itemdb-verify (repo scope) runs a builder's read-only --verify, which itemdb-build cannot pass.

Fixed -- gear school resistances on Vanilla, TBC and Wrath (MINOR 43, Dibs' DIBSREQ-IDB)

Dibs' contract (inbox f93b32b37ec4): a frost-resistance sort in Dibs' Planner found almost nothing on Classic Era, because GetStats returned no school resistance for most gear. Glacial Cloak (22658) had stamina and armor but not its 24 frost; Polar Tunic (22661) was missing 40 frost and Arcanist Robes (16798) 10 fire.

• Root cause: the Classic clients' GetItemStats never reports ItemSparse.Resistances_1..6 (0 items in the Vanilla and TBC walks, as build-core-wago.py already recorded), so the walk-built cores carried only resistance that comes from an on-equip aura (16061: column 0, walk fire 9). The Wrath core reproduces the TBC walk and so lacked them too.
• tools/build-equip-stats.py now reads Resistances_1..6 in its existing ItemSparse pass and emits RESISTANCE<n>_NAME (1 holy, 2 fire, 3 nature, 4 frost, 5 shadow, 6 arcane) into EquipStats.lua, which GetStats merges. An item whose core ALREADY carries a school (an on-equip aura the walk reported) and whose ItemSparse column names the same school would need the two summed, which GetStats' core-wins merge cannot do -- so the builder REFUSES and names the item (Peer Review 13334356 F1; the first cut only counted it on stderr). None exists on any version today.
• Rebuilt: Vanilla 1,371 items (pinned to 1.15.9.69722, the build its siblings carry, then split-seasonal.py), TBC 667, Wrath 723. Forever gains none: its ItemSparse columns are empty and its walk already reports resistances. Mists, Cata and Retail store resistances as stat slots, which their cores already read; not rebuilt.
• itemdb_common.core_rows is now THE reader of the shipped class files (CORE_ROW + the lib:LoadCore( test), and core_ids is a view over it. It replaces four private copies of the row regex and file walk in build-equip-stats.py, build-core-wago.py, build-core-wowsims.py and derive-core-maps.py (whose hand-kept NON_ITEM side-table list is gone). Found through it: core_resist_keys assigned per row, so an id with a base AND a seasonal row kept only the last file's schools (Peer Review 13334356 F1b); it now takes the union.
• Tests/data_baseline.json re-recorded. The 14 shrinks it accepted were all already in the released v1.4.0 tree, HEAD being that release: Data/Retail/_core/Effects.lua 1,186 -> 1,182 (-4), Data/Vanilla/_seasonal/Quest.lua 451 -> 450 (-1) and the twelve Data/Vanilla/_seasonal/<locale>/Names.lua 6,523 -> 6,522 (-1 each; the same item moved to the base Quest.lua / Names.lua, both +1). Also adopted: Retail Hidden.lua (new, 13,923) and the grown Hidden.lua on Cata, Forever, Mists, TBC and Wrath.
• MINOR 42 -> 43 so a consumer can gate on the data (select(2, LibStub("LibItemDB-1.0")) >= 43); no API change. CLAUDE.md's MINOR rule now says so: it moves for an API change OR when a consumer needs to detect new data (Peer Review 13334356 F2).
• Tests/equip_resist_check.py red-checked by removing Glacial Cloak's row from the Vanilla file: it failed on that item, and passed again once restored.

Older versions (v0.0.1 to v1.4.0) are in CHANGELOG_ARCHIVE.md.