Ability verification and lifecycle fixes

Scope and evidence

The September 2026 review at 41602e54 inspected all 39 Warcraft III game/skills/*.c files. Three agents covered combat spells, orders/economy, and passives/items/campaign handlers; the coordinator covered shared dispatch. The original baseline passed 1,132 WC3 tests in each archive mode, but sixteen new behavior scenarios failed and enabling Human autocast crashed. Registry presence and status-record creation had overstated working functionality.

The fixes retain these reproductions as permanent tests, extend their inverse and lifecycle coverage, and correct additional failures exposed by that work. The four new suites contain 37 tests / 302 assertions, passing in both ROC and TFT modes. Five save-schema tests additionally cover the new channel fields. This verifies the named contracts, not exhaustive retail parity or rendered art.

Fixed contracts

Area	Previous failure	Owning fix and regression evidence
Human autocast	A boolean message was dereferenced as a target pointer; aliases used base data	Human, validated-spell, Avatar and bolt procedures decode only the current message's union member. The command/HUD/scheduler retain the actual rawcode. Alias cost/heal, runtime membership, switching, removal and unsupported-policy rollback are tested.
Repair autocast	Switching related procedures could clear their shared policy flag	Disable the old policy before enabling its replacement; disabling a nonselected alias has no effect. Rejected enable restores the previous policy.
Hold Position	Out-of-range acquisition walked toward enemies	Attack returns to stationary scanning; an in-range enemy can still be attacked.
Delayed damage	An old projectile kill replaced a newer Move or consumed its queue	Combat completion requires the current Attack procedure and matching target. Animationless post-hit fallback also requires unchanged move and target.
Explicit Attack/Smart	Replaced Follow/Patrol/Attack-move resumed after combat	`S_OrderAttack` clears retained movement only after accepting an explicit order. Autonomous acquisition keeps continuation; queued orders replace it only when executed.
Repair completion	`building->stand` erased the producer queue	Repair owns the worker's completion only; the target keeps its production behavior and queue. Tests advance the retained trainee.
Flame Strike	Damage was treated as a long cast delay; pulses ran every frame	Read Cast/HeroDur/Dur and Hfs1–Hfs6 correctly; preserve absolute fractional pulse deadlines, phases, building multiplier and aggregate cap.
Blizzard	Zero Dur truncated six authored waves to two	Wave count and authored interval determine completion, independently of the zero-duration sentinel.
Life Drain / Siphon Mana	Life Drain affected mana; channels survived cancellation	Separate authored health/mana fields and allied transfer; validate each cast, tether and entity incarnation at every pulse.
Channel lifecycle	No-target casts lacked state; old thinkers survived recasts or stopped new orders	Start all channel shapes consistently; capture a cast serial and owner/target incarnation; route move-leave to the active procedure and explicitly cancel on Stop. Stun, movement, recast, expiry, reuse and save continuation are tested.
Aerial Shackles	The caster buff was applied to the victim, so movement/attacks remained allowed	Read the target member of the authored buff pair. Remove its lock when that channel ends, preserving a replacement cast's lock.
Storm Bolt	Spell immunity blocked damage but not stun; creep level selected Hero duration	Check impact acceptance before applying stun and use actual Hero identity for HeroDur.
Frost Nova	A no-target burst damaged around the caster	Register a unit target; apply direct DataB damage and area DataA damage around that target.
Resurrection	Only Heroes were eligible	The no-target spell finds ordinary friendly corpses in the authored area/count, reuses their edicts, retires death state and restores unit/food activity. Reject an empty cast before committing resources.
Attribute items	Strength, Agility and Intelligence were rotated	Both permanent tomes and passive attribute aliases use DataA=Agility, DataB=Intelligence, DataC=Strength.
Passive items	Exact base-code comparisons discarded stock alias bonuses	`A_ITEM_ADD` / `A_ITEM_REMOVE` carry `abilityitem_t` to the owning passive procedure. Read each alias's amount on pickup/drop; no per-family global amount cache. Stacked items survive save/load and independent removal.
Natural creep sleep	Wake removed unrelated owned overlays	Tag sleep art with ACsp and require that identity during cleanup. Wake, move and disable preserve regeneration art.

Retail target validation checks

Two focused JASS probes checked Charm (ANch) and TFT Purge (Apg2) against the pinned Warcraft III 1.29.2.9231 executable (PE32 x86; SHA-256 3f2ed0120d80578bf07e4423296dade1adfb959d59a2d20a7584224559570eed).

Charm transferred an ordinary level-3 nftt creep. A Hpal Hero, a dead nftt, and a live nftt with permanent Resistant Skin (ACsk) remained with Player 12. Follow-up tests found the stock level-5 boundary: levels 4 and 5 transferred; level 6 did not. A custom level-3 copy behaved anomalously and remains unexplained.
Friendly and hostile Purge casts both spent mana. The allied target moved about as far as its unpurged control at 1.5 seconds (382.881 vs. 403.491); the hostile target was strongly suppressed (80.865 vs. 385.253). This supports applying Purge's movement status to hostile targets only while retaining friendly dispel behavior. Exact recovery and other aliases remain unverified.

These are Retail JASS observations, not a complete target-mask specification. The implementation rejects dead or resistant Charm targets and gates Purge's movement status on hostility. Regression tests cover these exclusions and relationship-dependent Purge effects in games/warcraft-3/game/tests/t_charm.c and t_purge.c. The complete classic/TFT WC3 suites run through make test.

Data and timing

Use the archive audit instead of inferring fields from tooltip prose:

build/bin/ability_audit -data 'data/Warcraft III' -roc -raw AIsm
build/bin/ability_audit -data 'data/Warcraft III' -tft -raw AHfs
python3 tools/wc3_ability_class_audit.py --format=coverage

The active TFT Flame Strike row uses Hfs1/DataA for full pulse damage, Hfs2/DataB for its interval, Hfs3/DataC for lesser pulse damage, Hfs4/DataD for its interval, Hfs5/DataE for the building multiplier, and Hfs6/DataF for the aggregate damage cap. Cast is the initial delay; HeroDur supplies the full-damage phase and Dur the damage lifetime. Deadlines advance from the previous scheduled pulse, so a 0.33-second interval does not become 0.4 seconds on every 100 ms server tick. Positive sub-frame intervals are caught up at the next simulation step; invalid intervals are rejected/logged.

Both ROC and TFT archive audits confirm AItg → AIat (attack bonus 1), AId1 → AIde (armor 1), and the three tome attribute columns. Combat fixtures include an explicit ROC numbered Data11/Data12 conversion test for Blizzard. Running TFT numeric fixtures in ROC mode verifies runtime compatibility; it does not independently verify every ROC authored value or TFT-only spell.

Dispatch and persistence

abilityCall_t is a message-tagged union. A_COMMAND supplies a client, A_INIT a classname, A_AUTOCAST_SET a boolean, and only target-bearing messages supply a spellTarget_t. Never eagerly read the target union member before deciding which message is being handled.

edict.channel.code is nonzero on the caster only. Each channel thinker retains its spell rawcode in class_id, the cast channel.serial, and owner_spawn_time / target_spawn_time where applicable. Reused owner/target slots and replacement casts cannot inherit an old effect. S_SpellEndChannel clears only a matching cast, then releases its thinker; cancellation does not install Stand over an already accepted replacement order.

Save format 21 adds this channel identity state to the edict schema. Older save versions are rejected. The new channel callbacks are appended to save_cfunctions[]; never reorder existing callback indexes. The saved channel origin is preserved for movement cancellation after loading. A live drain and stacked passive item aliases have production WriteGame/ReadGame continuation tests, in addition to the individual schema tests.

Tests

make test-wc3-engine WC3_PATTERN='wc3_ability_*'
make test-wc3-engine WC3_PATTERN='wc3_order_lifecycle.*'
make test-wc3-engine WC3_PATTERN='wc3_item_lifecycle.*'
make test-wc3-engine WC3_PATTERN='wc3_save.*'
make test
make build

Permanent regression sources are game/tests/t_ability_dispatch.c, t_ability_lifecycle.c, t_order_lifecycle.c, and t_item_lifecycle.c. Existing combat, building, spell and save tests also check affected paths. Final verification on 2026-09-13: make test passes 3,228 test invocations / 65,585 assertions across 20 suite invocations, including 1,174 WC3 tests / 25,316 assertions per archive mode. make build passes. Both builds emit no compiler warnings; registry coverage remains 193 registered TFT class IDs. Use full local socket access for the aggregate suite's UDP tests; sandbox bind denial is not a gameplay failure. No interactive game launch was needed.

Limits and history

The fixes address reproduced defects and their related lifecycle cases. Previously documented partial implementations remain partial: Wisp/Acolyte harvesting, mine transformations, general transport drop, Purchase Item, Mirror Image's complete illusion semantics, several status-only spells, and Siphon Mana's temporary above-maximum mana/decay behavior. Visual effects, full target-mask vocabulary and comprehensive retail spell parity remain separate work. Do not interpret a registered procedure as complete behavior.

Blame traced the unsafe Human union reads to e12b63ad0, Flame/Drain field and timer errors to 22ff3692, item alias/attribute errors to 85fd0906, attack continuation to 317fc09ae, Repair completion to 6d5025e92, and sleep overlay cleanup to ebbe30e2. These histories explain why the new regressions exercise caller/event order, authored aliases and inverse behavior rather than only testing procedure lookup.

See ability coverage, ability implementation, save/load, and autocast.

Jaina / Archmage follow-up (September 19, 2026)

The Archmage review tightened three abilities without adding Hero-specific branches. AHwe now treats DataA as the authored summon count, records the creating ability on each summon, and emits TargetArt while retaining the shared timed-life/event path. AHbz now evaluates its own authored target mask, applies DataD to structures after the DataF per-wave cap, and emits DataC EffectArt shards without using the visual positions for hit detection.

AHmt now participates in the shared channel lifecycle instead of relocating immediately. DataB owns the completion delay; a thinker retains the destination entity incarnation, temporarily pauses that destination, owns persistent area effects, and cleans all of those resources on success or cancellation. Completion uses the current destination position, keeps the caster inside the DataA total, excludes allied-player armies and structures from the payload, honors DataC clustering, and uses ordinary unstuck placement. The thinker callback is appended to the save callback roster, and regression coverage includes delayed completion, unit-cap/ownership filtering, structure destinations, destination death, cancellation cleanup, Water Elemental DataA ownership, Blizzard structure scaling, and production save/load continuation.

Brilliance Aura remains on the existing cached Hero-aura path. This follow-up does not replace that path with the newer generic target-token helper because the current unit targtype representation conflates movement class and organic/mechanical classification; doing so here could regress stock mechanical recipients. That broader target-model cleanup remains separate work.

Jaina / Archmage remaining-fidelity follow-up (September 19, 2026)

A second conservative pass closes behavior that the bundled Warsmash source makes explicit without broadening the shared damage/type model. Blizzard now waits for its authored Cast interval before the first shard wave, then splits each wave into shard presentation followed by damage 0.8 seconds later. The phase is stored on the already-serialized thinker state, so saving between the shards and impact resumes the pending damage phase rather than replaying art. The existing target mask, DataD building multiplier and DataF aggregate cap remain authoritative. Cold/spell damage typing and spell-sound parity are still separate shared-system work.

Brilliance Aura now consumes the stock air,ground,friend,self target mask in the cached Hero-aura path rather than treating every friendly entity in range as a recipient. Rows that omit targs keep the historical friendly fallback for sparse ROC/custom fixtures. This deliberately does not rewrite the engine-wide organic/mechanical classification model; that broader cleanup is still separate.

Water Elemental keeps its existing DataA/UnitID/timed-life path but now preserves the caster facing and relinks collision-displaced summons. Mass Teleport likewise relinks every relocated unit (and dirties FOW blocker state where relevant), so the server broad phase no longer retains the source position after the visual and simulation coordinates have moved. Regression coverage documents Blizzard's 0.8-second phase timing and save/load continuation, Brilliance target filtering, Water Elemental separation/facing, and Mass Teleport vacated-space relinking.

Jaina / Archmage data-driven aura and targeting follow-up (September 19, 2026)

A third conservative pass closes two data-driven gaps that are explicit in the bundled Warsmash implementation. Brilliance Aura now treats DataB as its flat-versus-percentage switch: flat DataA remains MP/sec, while percentage DataA scales the recipient's intrinsic mana regeneration (base unit regeneration, ordinary mana-regeneration bonuses, and the current Intelligence regeneration term). Flat and percentage Brilliance providers keep separate strongest-copy slots, matching the separate non-stacking stat families used by Warsmash. The Hero-aura cache also resolves the actual authored alias and rank before reading Area, targs, DataA, and DataB, so custom Brilliance-derived abilities no longer silently fall back to the stock AHab row.

The shared unit-target validator now reads targs from the caster's current ability rank instead of always using level 1. This matters to Blizzard and Mass Teleport when custom object data changes target classes across ranks, and it preserves the requested alias as the data source. Focused regressions cover rank-specific target masks and a custom percentage-mode Brilliance alias mixed with a stock flat provider.

This pass still does not invent cold/spell damage typing, per-shard sound semantics, or an engine-wide organic/mechanical target-class rewrite. Those remain shared-system fidelity work because the current OpenRealm representations do not model them cleanly enough for a low-risk Jaina-specific change.

Jaina / Archmage summon identity and Blizzard effect-object follow-up (September 19, 2026)

A fourth conservative pass closes two presentation/lifecycle details that the bundled Warsmash implementation makes explicit without extending OpenRealm's shared damage model. Summon Water Elemental now also applies the ability's authored BuffID to each created Elemental. OpenRealm deliberately keeps BTLF as the authoritative timed-life clock because the existing timed-life bar and UnitPauseTimedLife native are built around that status; the authored Water Elemental buff supplies the WC3 ability-specific status/presentation identity without introducing a second expiration timer. BTLF is now treated as non-dispellable lifecycle state by the shared Dispel/Purge filter, so dispel still deals its authored damage to summoned units but cannot accidentally turn a surviving timed summon into a permanent unit.

Blizzard now resolves each shard's EffectArt through the level's authored EfctID object, matching Warsmash's effectId path, instead of resolving the visual directly from the casting ability alias. Focused regressions cover the Water Elemental BuffID plus retained BTLF, preservation of timed life after Dispel Magic, and the Blizzard EfctID art lookup.

That fourth pass intentionally left cold/spell damage typing and per-shard ability sound alone because both required shared-system support. The following pass adds the sound-data path; typed SPELLS/COLD damage still remains separate.

Jaina / Archmage ability-sound follow-up (September 19, 2026)

A fifth conservative pass closes Blizzard's remaining high-confidence sound-data gap without changing spell damage semantics. The typed metadata registry now loads UI\SoundInfo\AbilitySounds.slk with the same sound-row schema already used for unit/UI sounds, and ability/buff presentation exposes authored Effectsound / Effectsoundlooped fields. G_PlayAbilityEffectSound resolves an alias from the requested object or its base object and emits a positional world sound at the requested effect point using the authored volume.

Blizzard now uses the level EfctID for both shard art and shard sound, matching the bundled Warsmash implementation's spawnSpellEffectOnPoint(effectId) plus unitSoundEffectEvent(caster, effectId) sequence. The sound fires once for each DataC shard during the presentation phase, never during the delayed damage phase. Regression coverage injects a synthetic AbilitySounds.slk row and verifies six shard sounds, alias/path resolution, positioned playback, and 0-127 volume scaling. The test MPQ also carries a minimal AbilitySounds fixture and verifies it is packed.

The server intentionally selects the first authored ability-sound file rather than calling rand(): Warsmash performs variant selection in rendering/audio code, while using OpenRealm's simulation RNG for presentation would perturb deterministic game state. Client-side random variant selection plus authored pitch/pitch-variance are still presentation refinements. SPELLS/COLD Blizzard damage typing, the broader organic/mechanical target model, and finer Mass Teleport placement/order parity remain separate shared-system work.