Derived durations, and the measurement that was one tick long

[JDS300]

Cold-start countdowns for DoTs and debuffs, drawn from the wiki's base duration scaled by spell rank and marked as estimates — plus a fix to the measurement that reported every DoT one tick too long.

The anchoring task changed direction

The brief asked to re-anchor DoT timers on SpellCastEvent. The log says no. Six DoTs with both fade lines and an exact wiki duration, medians over 138 completed casts:

Spell	wiki	tick1→fade	cast→fade	old +6
Immolate	48	48 ✓	53	54
Drones of Doom	48	48 ✓	53	54
Gasping Embrace	48	48 ✓	54	54
Stinging Swarm	54	54 ✓	59	60
Vengeance of the Wild	30	31 ✓	38	37
Drifting Death	1 minute	54	59	60

Anchoring on the cast matches nothing — it runs long by each spell's own cast time (Immolate 2.5s, Shiftless Deeds 6.0s), which is why the error was never a constant six seconds. The first tick is the landing, and the fade line arrives at the last tick, not one after it. The real defect was a phantom trailing tick in the no-fade measurement path. Drifting Death is the lone miss and explains itself: its wiki value is the prose 1 minute, rounded from 54.

Three tiers of trust

Measured (from the log, authoritative) > Derived (base × (1 + 0.10 × tier), shown ~2:40) > Unknown (--, never a guessed number). A measurement is never corrected toward the wiki: Tepid Deeds keeps its measured 126s against a wiki 150 whose own page contradicts itself.

Base durations are shipped, not fetched

Data/SpellDurations.json (680 entries) is promoted from the harvest already in the repo by scripts/harvests/eqlwiki/spells-promote.py. No HTTP mid-fight, works offline. Instant and level-scaled durations (Cripple's 6.3 minutes @L53 to 7.0 minutes @L60) are excluded — an absent spell means unknown, never a guess.

The catalog is the authority on names

121 harvested spells are genuinely named with a trailing numeral — Clarity II, Burnout IV — and are not ranks. So an exact catalog hit beats a rank interpretation. Read the other way round, Burnout IV would resolve to 3600 × 1.4 = 5040s instead of 900s.

Validation against the real 690k-line log

• Shiftless Deeds IV measured 214.0s against a derived 210.0s — the ×1.1/tier formula reproducing itself from an independent direction.
• Cold-start handover works: Shiftless Deeds IV showed 11,922 derived samples, then 62,129 measured. Tepid Deeds 37,928 → 31,083.
• Heroic Leap I: 7,987 samples, always --. No catalog entry, no invented number.
• Real app, isolated profile: Drifting Death --, Immolate 0:00, Tepid Deeds ~1:20, Shiftless Deeds IV ~2:18.

Bugs found and fixed along the way

• _recastPending stored the ranked cast name and was read with the unranked tick name, so recast detection could never fire for a ranked DoT — the exact failure its own comment documents. Every key now goes through BaseNameOf.
• A derived expiry retired a slow before its fade arrived, so the spell could never graduate to Measured — pinned at the estimate forever. Shiftless Deeds IV escaped by one second. Retirement and forgetting are now separate concerns (_awaitingFade).
• A long derived duration defeated UnknownCap, holding a chip for 54 minutes.
• The promote script resolved 13 name collisions silently; two differ materially (Rabies 2880 vs 314). It now warns instead of choosing in silence.

Tests

EQBuddy.Tests 842 → 882, EQBuddy.Avalonia.Tests 81. All green.

Known, out of scope

The slice-2 landing/cast pairing mis-attributes: Stun, Wrath, Valor and others surface as slow chips because the landing line names no spell. Pre-existing, but this feature upgrades two of them from -- to a confident wrong countdown. The harvest carries a beneficial flag and a beneficial spell can never explain a "slows down" line — worth its own slice.

🤖 Generated with Claude Code